SC117/Swift-1.5-Qwen3.8-Flash-Next-GSQ-RCO-abliterated-GGUF overview

---

base_model: ukisai/Swift-Qwen3.8-Flash-Next

base_model_relation: quantized

license: other

license_name: swift-open-license-1.0

license_link: https://huggingface.co/ukisai/Swift-1.5-Qwen3.8-Flash-Next-GSQ-RCO-GGUF/blob/main/LICENSE

library_name: gguf

pipeline_tag: image-text-to-text

language:

- en

- zh

tags:

- gguf

- llama-cpp

- abliterated

- uncensored

- quantization

- tensor-transplant

- gsq

- rco

- moe

- mixture-of-experts

- qwen3.8

- flash-next

- long-context

- reasoning

- reasoning-efficient

- token-efficient

- red-teaming

---

<div style="font-family: -apple-system, BlinkMacSystemFont, 'Segoe UI', Roboto, sans-serif; margin-bottom: 24px;">

<div style="background: #0f172a; border-radius: 20px; padding: 40px 32px 30px 32px; text-align: center; position: relative; overflow: hidden;">

<div style="position: absolute; top: -40px; right: -40px; width: 160px; height: 160px; background: rgba(139,92,246,0.18); border-radius: 50%;"></div>

<div style="position: absolute; bottom: -50px; left: -30px; width: 140px; height: 140px; background: rgba(59,130,246,0.15); border-radius: 50%;"></div>

<div style="position: absolute; top: 30%; right: 12%; width: 56px; height: 56px; background: rgba(59,130,246,0.22); border-radius: 50%;"></div>

<div style="display: inline-flex; flex-wrap: wrap; justify-content: center; gap: 8px; margin-bottom: 18px; position: relative; z-index: 1;"><span style="background: #10b981; color: white; font-size: 11px; font-weight: 700; padding: 5px 14px; border-radius: 20px;">Swift 1.5 upstream · layout preserved</span><span style="background: #3b82f6; color: white; font-size: 11px; font-weight: 700; padding: 5px 14px; border-radius: 20px;">144 tensors transplanted</span><span style="background: #8b5cf6; color: white; font-size: 11px; font-weight: 700; padding: 5px 14px; border-radius: 20px;">IQ3_S / IQ3_XXS / IQ2_XS / Q2_0</span><span style="background: #f59e0b; color: #0f172a; font-size: 11px; font-weight: 700; padding: 5px 14px; border-radius: 20px;">per-tensor blake2b verified</span><span style="background: #0ea5e9; color: white; font-size: 11px; font-weight: 700; padding: 5px 14px; border-radius: 20px;">reasoning efficiency kept</span><span style="background: #1e293b; color: #94a3b8; font-size: 11px; font-weight: 700; padding: 5px 14px; border-radius: 20px;">Swift Open License</span></div>

<h1 style="margin: 0 0 10px 0; font-size: 34px; font-weight: 800; color: #f8fafc; letter-spacing: -0.5px; border: none; position: relative; z-index: 1;">Swift-1.5-Qwen3.8-Flash-Next · GSQ-RCO-abliterated</h1>

<p style="margin: 0 0 20px 0; font-size: 15px; color: #cbd5e1; position: relative; z-index: 1;">The refusal direction removed from UkisAI's <b>reasoning-efficient</b> derivative of Qwen3.8-Flash-Next. 144 tensors across all 48 layers swap in ready-made abliterated weights by <b>byte-level transplant</b> — every GSQ value and scale is left untouched, the file size moves by <b>less than 0.2%</b>, and the <b>reasoning-efficiency of Swift 1.5 survives</b> (still 34.8% below the base).</p>

<div style="position: relative; z-index: 1; margin: 0 auto; width: 72%; max-width: 540px;">

<div style="height: 8px; border-radius: 8px; background: #334155; position: relative; overflow: visible;">

<div style="height: 8px; width: 100%; border-radius: 8px; background: linear-gradient(90deg, #3b82f6, #8b5cf6);"></div>

<div style="position: absolute; top: 50%; left: 0%; transform: translate(-50%, -50%); width: 18px; height: 18px; border-radius: 50%; background: #ffffff; box-shadow: 0 0 0 4px rgba(59,130,246,0.35);"></div>

<div style="position: absolute; top: 50%; left: 100%; transform: translate(-50%, -50%); width: 18px; height: 18px; border-radius: 50%; background: #ffffff; box-shadow: 0 0 0 4px rgba(139,92,246,0.35);"></div>

</div>

<div style="display: flex; justify-content: space-between; margin-top: 8px; font-size: 10px; color: #94a3b8; font-weight: 600;"><span>61.9 GB · Q2_0</span><span style="color: #c4b5fd;">four tiers all within 0.2% of upstream</span><span>77.8 GB · IQ3_S</span></div>

</div>

<p style="margin: 18px 0 0 0; font-size: 13px; position: relative; z-index: 1;"><span style="color: #94a3b8;">English</span><span style="color: #475569;"> · </span><a href="https://huggingface.co/SC117/Swift-1.5-Qwen3.8-Flash-Next-GSQ-RCO-abliterated-GGUF/blob/main/README.zh-CN.md" style="color: #93c5fd; text-decoration: none; font-weight: 600;">简体中文</a></p>

</div>

</div>

<div style="font-family: -apple-system, BlinkMacSystemFont, 'Segoe UI', Roboto, sans-serif; display: flex; flex-direction: column; gap: 18px; margin-bottom: 30px;">

<div style="border: 1px solid #cbd5e1; border-radius: 12px; overflow: hidden; background: #ffffff; box-shadow: 0 2px 4px rgba(0,0,0,0.02);"><div style="background: linear-gradient(135deg, #2563eb 0%, #7c3aed 100%); padding: 12px 16px; color: white; font-weight: 700; font-size: 14px; display: flex; align-items: center; gap: 8px;"><span>🧭</span> What this is</div><div style="padding: 16px; font-size: 13px; color: #334155; line-height: 1.7;"><p style="margin: 0 0 10px 0;">The <b>abliterated edition</b> of UkisAI's <b>Swift 1.5</b> — itself a reasoning-efficient derivative of Qwen3.8-Flash-Next that uses ~52% fewer thinking tokens on coding benchmarks. Every GSQ-learned quantized tensor is left bit-for-bit untouched; only the 144 tensors that write back into the residual stream are replaced, layer by layer, with the corresponding tensors from a ready-made abliterated release.</p><table style="width: 100%; border-collapse: collapse; font-size: 12.5px;"><tbody><tr><td style="padding: 6px 10px; ">Base</td><td style="padding: 6px 10px; "><a href="https://huggingface.co/ukisai/Swift-1.5-Qwen3.8-Flash-Next-GSQ-RCO-GGUF" style="color: #1d4ed8; text-decoration: none; font-weight: 600;">ukisai/Swift-1.5-Qwen3.8-Flash-Next-GSQ-RCO-GGUF</a> —— GSQ learned quantization + RCO budget allocation, plus UkisAI&rsquo;s anti-overthinking post-training</td></tr><tr><td style="padding: 6px 10px; box-shadow: 0 1px 0 0 rgba(128,128,128,0.15); ">Abliterated weights</td><td style="padding: 6px 10px; box-shadow: 0 1px 0 0 rgba(128,128,128,0.15); "><a href="https://huggingface.co/orcarouter/Qwen3.8-Flash-Next-Uncensored-GGUF" style="color: #1d4ed8; text-decoration: none; font-weight: 600;">orcarouter/Qwen3.8-Flash-Next-Uncensored-GGUF</a> —— the 144 target tensors&rsquo; bytes come straight from here</td></tr><tr><td style="padding: 6px 10px; box-shadow: 0 1px 0 0 rgba(128,128,128,0.15); ">Method</td><td style="padding: 6px 10px; box-shadow: 0 1px 0 0 rgba(128,128,128,0.15); ">GGUF &rarr; GGUF <b>byte-level tensor transplant</b>. Only where the official type demands it (the 48 routed-expert <code>ffn_down_exps</code> stored as Q2_0) are the abliterated values re-encoded — using <b>Swift&rsquo;s own learned block scales</b>, so no GSQ value is recomputed, only the code words change.</td></tr></tbody></table></div></div>

<div style="border: 1px solid #fca5a5; border-radius: 12px; overflow: hidden; background: #ffffff; box-shadow: 0 2px 4px rgba(0,0,0,0.02);"><div style="background: linear-gradient(135deg, #dc2626 0%, #ea580c 100%); padding: 12px 16px; color: white; font-weight: 700; font-size: 14px; display: flex; align-items: center; gap: 8px;"><span>⚠️</span> Safety notice</div><div style="padding: 16px; font-size: 13px; color: #334155; line-height: 1.7;"><p style="margin: 0 0 8px 0;">This model is <b>abliterated</b> — the refusal direction has been removed and it has <b>no reliable built-in guardrails</b>. It may comply with harmful, illegal or unsafe requests. Deploy it only where you can supply your own moderation, access control and legal review; do not put it in front of end users without your own safety layer.</p><p style="margin: 0; color: #64748b;">Also note the upstream <b>Swift Open License v1.0</b> is stricter than Apache-2.0: commercial use is free only below US$1M annual revenue.</p></div></div>

<div style="border: 1px solid #c4b5fd; border-radius: 12px; overflow: hidden; background: #ffffff; box-shadow: 0 2px 4px rgba(0,0,0,0.02);"><div style="background: linear-gradient(135deg, #7c3aed 0%, #a855f7 100%); padding: 12px 16px; color: white; font-weight: 700; font-size: 14px; display: flex; align-items: center; gap: 8px;"><span>🧬</span> Why a transplant and not a requantization</div><div style="padding: 16px; font-size: 13px; color: #334155; line-height: 1.7;"><p style="margin: 0 0 10px 0;"><b>GSQ</b> treats each tensor&rsquo;s grid and scales as learnable parameters, trained jointly with a Gumbel-Softmax relaxation; <b>RCO</b> then allocates a quantization type per tensor under a global size budget. A GSQ-RCO file therefore does <b>not</b> equal "BF16 weights pushed through a standard quantizer" — re-running the recipe yields a <b>different</b> set of values.</p><p style="margin: 0 0 12px 0; padding: 10px 14px; background: #f5f3ff; border-left: 4px solid #7c3aed; border-radius: 6px; color: #4c1d95; font-weight: 600;">There is exactly one way to abliterate a GSQ-RCO file: <b>replace the tensors in place, inside the already-quantized GGUF</b>. That is what this repo does.</p><p style="margin: 0 0 10px 0;">On top of that, Swift 1.5 is <b>not</b> the plain base — its post-training is what makes it reason efficiently. That behaviour lives in the very tensors being replaced, so the only honest way to check the transplant is to <b>measure it</b> afterwards (see the cost table below).</p><p style="margin: 0; font-size: 12.5px; color: #64748b;">The ablation itself is a strictly rank-1 linear edit — <code>W &larr; W &minus; r(r&#7488;W)</code> — sharing one direction <code>r</code> across all 48 layers, both attention types, both sites and all 512 routed experts, so "copy a ready-made ablated tensor" and "compute the ablation yourself and quantize" lie on the same line — and the former adds no new quantization error.</p></div></div>

<div style="border: 1px solid #cbd5e1; border-radius: 12px; overflow: hidden; background: #ffffff; box-shadow: 0 2px 4px rgba(0,0,0,0.02);"><div style="background: linear-gradient(135deg, #2563eb 0%, #7c3aed 100%); padding: 12px 16px; color: white; font-weight: 700; font-size: 14px; display: flex; align-items: center; gap: 8px;"><span>📦</span> Files in this repo</div><div style="padding: 16px; font-size: 13px; color: #334155; line-height: 1.7;"><p style="margin: 0 0 12px 0;">Two shards per tier. <b>Unlike the ISTA GSQ-RCO family, Swift&rsquo;s shard 2 is not a shared n-gram table</b> — it holds <b>layers 13–47</b>, i.e. real quantized weights, so every tier needs its own shard 2. Shard 1 looks nearly the same size across tiers only because the 28.8 GB <code>per_layer_token_embd</code> table (72% of shard 1) is tier-independent and is not touched by the ablation.</p><table style="width: 100%; border-collapse: collapse; font-size: 12.5px;"><thead><tr style="background: rgba(37,99,235,0.05);"><th style="padding: 7px 10px; border-bottom: 2px solid #2563eb; text-align: left; color: #1d4ed8; ">Folder</th><th style="padding: 7px 10px; border-bottom: 2px solid #2563eb; text-align: left; color: #1d4ed8; ">Shard 1 (weights)</th><th style="padding: 7px 10px; border-bottom: 2px solid #2563eb; text-align: left; color: #1d4ed8; ">Shard 2 (layers 13–47)</th><th style="padding: 7px 10px; border-bottom: 2px solid #2563eb; text-align: left; color: #1d4ed8; ">Total</th><th style="padding: 7px 10px; border-bottom: 2px solid #2563eb; text-align: left; color: #1d4ed8; ">vs upstream</th><th style="padding: 7px 10px; border-bottom: 2px solid #2563eb; text-align: left; color: #1d4ed8; ">Status</th></tr></thead><tbody><tr><td style="padding: 6px 10px; "><span style="font-family: monospace; font-weight: 700; color: #1d4ed8;">IQ3_S/</span></td><td style="padding: 6px 10px; "><b>41.77 GB</b></td><td style="padding: 6px 10px; ">36.04 GB</td><td style="padding: 6px 10px; "><b>77.81 GB</b></td><td style="padding: 6px 10px; "><span style="color: #92400e; font-weight: 700;">−0.18 GB</span></td><td style="padding: 6px 10px; "><span style="color: #166534; font-weight: 700;">✅ built · verified</span></td></tr><tr><td style="padding: 6px 10px; box-shadow: 0 1px 0 0 rgba(128,128,128,0.15); "><span style="font-family: monospace; font-weight: 700; color: #1d4ed8;">IQ3_XXS/</span></td><td style="padding: 6px 10px; box-shadow: 0 1px 0 0 rgba(128,128,128,0.15); "><b>37.04 GB</b></td><td style="padding: 6px 10px; box-shadow: 0 1px 0 0 rgba(128,128,128,0.15); ">33.58 GB</td><td style="padding: 6px 10px; box-shadow: 0 1px 0 0 rgba(128,128,128,0.15); "><b>70.62 GB</b></td><td style="padding: 6px 10px; box-shadow: 0 1px 0 0 rgba(128,128,128,0.15); "><span style="color: #92400e; font-weight: 700;">−0.13 GB</span></td><td style="padding: 6px 10px; box-shadow: 0 1px 0 0 rgba(128,128,128,0.15); "><span style="color: #166534; font-weight: 700;">✅ built · verified</span></td></tr><tr><td style="padding: 6px 10px; box-shadow: 0 1px 0 0 rgba(128,128,128,0.15); "><span style="font-family: monospace; font-weight: 700; color: #1d4ed8;">IQ2_XS/</span></td><td style="padding: 6px 10px; box-shadow: 0 1px 0 0 rgba(128,128,128,0.15); "><b>37.06 GB</b></td><td style="padding: 6px 10px; box-shadow: 0 1px 0 0 rgba(128,128,128,0.15); ">26.41 GB</td><td style="padding: 6px 10px; box-shadow: 0 1px 0 0 rgba(128,128,128,0.15); "><b>63.47 GB</b></td><td style="padding: 6px 10px; box-shadow: 0 1px 0 0 rgba(128,128,128,0.15); "><span style="color: #92400e; font-weight: 700;">−0.01 GB</span></td><td style="padding: 6px 10px; box-shadow: 0 1px 0 0 rgba(128,128,128,0.15); "><span style="color: #166534; font-weight: 700;">✅ built · verified</span></td></tr><tr><td style="padding: 6px 10px; box-shadow: 0 1px 0 0 rgba(128,128,128,0.15); "><span style="font-family: monospace; font-weight: 700; color: #1d4ed8;">Q2_0/</span></td><td style="padding: 6px 10px; box-shadow: 0 1px 0 0 rgba(128,128,128,0.15); "><b>37.06 GB</b></td><td style="padding: 6px 10px; box-shadow: 0 1px 0 0 rgba(128,128,128,0.15); ">24.88 GB</td><td style="padding: 6px 10px; box-shadow: 0 1px 0 0 rgba(128,128,128,0.15); "><b>61.94 GB</b></td><td style="padding: 6px 10px; box-shadow: 0 1px 0 0 rgba(128,128,128,0.15); "><span style="color: #92400e; font-weight: 700;">−0.04 GB</span></td><td style="padding: 6px 10px; box-shadow: 0 1px 0 0 rgba(128,128,128,0.15); "><span style="color: #166534; font-weight: 700;">✅ built · verified</span></td></tr><tr><td style="padding: 6px 10px; box-shadow: 0 1px 0 0 rgba(128,128,128,0.15); "><code>mmproj-*.gguf</code></td><td style="padding: 6px 10px; box-shadow: 0 1px 0 0 rgba(128,128,128,0.15); ">—</td><td style="padding: 6px 10px; box-shadow: 0 1px 0 0 rgba(128,128,128,0.15); ">—</td><td style="padding: 6px 10px; box-shadow: 0 1px 0 0 rgba(128,128,128,0.15); "><b>0.91 GB</b></td><td style="padding: 6px 10px; box-shadow: 0 1px 0 0 rgba(128,128,128,0.15); ">vision projector, shared with the base</td><td style="padding: 6px 10px; box-shadow: 0 1px 0 0 rgba(128,128,128,0.15); ">✅ included</td></tr><tr><td style="padding: 6px 10px; box-shadow: 0 1px 0 0 rgba(128,128,128,0.15); "><code>mtp-q2_0.gguf</code> + <code>strata/rt/</code></td><td style="padding: 6px 10px; box-shadow: 0 1px 0 0 rgba(128,128,128,0.15); ">—</td><td style="padding: 6px 10px; box-shadow: 0 1px 0 0 rgba(128,128,128,0.15); ">—</td><td style="padding: 6px 10px; box-shadow: 0 1px 0 0 rgba(128,128,128,0.15); "><b>1.57 GB</b></td><td style="padding: 6px 10px; box-shadow: 0 1px 0 0 rgba(128,128,128,0.15); ">Strata-only draft runtime</td><td style="padding: 6px 10px; box-shadow: 0 1px 0 0 rgba(128,128,128,0.15); ">✅ included</td></tr><tr><td style="padding: 6px 10px; box-shadow: 0 1px 0 0 rgba(128,128,128,0.15); "><code>mtp-Qwen3.8-Flash-Next-Q8_0.gguf</code></td><td style="padding: 6px 10px; box-shadow: 0 1px 0 0 rgba(128,128,128,0.15); ">—</td><td style="padding: 6px 10px; box-shadow: 0 1px 0 0 rgba(128,128,128,0.15); ">—</td><td style="padding: 6px 10px; box-shadow: 0 1px 0 0 rgba(128,128,128,0.15); "><b>4.14 GB</b></td><td style="padding: 6px 10px; box-shadow: 0 1px 0 0 rgba(128,128,128,0.15); ">draft head for llama.cpp</td><td style="padding: 6px 10px; box-shadow: 0 1px 0 0 rgba(128,128,128,0.15); ">✅ included</td></tr></tbody></table><p style="margin: 10px 0 0 0; padding: 9px 13px; background: #f0fdf4; border-left: 4px solid #16a34a; border-radius: 6px; color: #14532d;">✅ <b>The ablated build is <i>smaller</i> than upstream</b> on every tier — re-encoding the expert tensors at Q2_0 with Swift&rsquo;s own scales actually saves a few megabytes.</p></div></div>

<div style="border: 1px solid #cbd5e1; border-radius: 12px; overflow: hidden; background: #ffffff; box-shadow: 0 2px 4px rgba(0,0,0,0.02);"><div style="background: linear-gradient(135deg, #2563eb 0%, #7c3aed 100%); padding: 12px 16px; color: white; font-weight: 700; font-size: 14px; display: flex; align-items: center; gap: 8px;"><span>🔬</span> What exactly changed (all four tiers)</div><div style="padding: 16px; font-size: 13px; color: #334155; line-height: 1.7;"><p style="margin: 0 0 12px 0;">Of the 1224 tensors in each file, <b>144</b> were replaced — all in the "writes back into the residual stream" families, covering every layer:</p><table style="width: 100%; border-collapse: collapse; font-size: 12.5px;"><thead><tr style="background: rgba(37,99,235,0.05);"><th style="padding: 7px 10px; border-bottom: 2px solid #2563eb; text-align: left; color: #1d4ed8; ">Tensor family</th><th style="padding: 7px 10px; border-bottom: 2px solid #2563eb; text-align: left; color: #1d4ed8; ">Count</th><th style="padding: 7px 10px; border-bottom: 2px solid #2563eb; text-align: left; color: #1d4ed8; ">Upstream type</th><th style="padding: 7px 10px; border-bottom: 2px solid #2563eb; text-align: left; color: #1d4ed8; ">This repo</th><th style="padding: 7px 10px; border-bottom: 2px solid #2563eb; text-align: left; color: #1d4ed8; ">Params</th></tr></thead><tbody><tr><td style="padding: 6px 10px; "><span style="font-family: monospace; font-size: 11.5px;">blk..ssm_out.weight</span></td><td style="padding: 6px 10px; ">36</td><td style="padding: 6px 10px; "><span style="color: #64748b;">IQ4_XS×29 / IQ3_S×6 / Q6_K×1</span></td><td style="padding: 6px 10px; "><b>IQ4_XS</b></td><td style="padding: 6px 10px; ">566 M</td></tr><tr><td style="padding: 6px 10px; box-shadow: 0 1px 0 0 rgba(128,128,128,0.15); "><span style="font-family: monospace; font-size: 11.5px;">blk..attn_output.weight</span></td><td style="padding: 6px 10px; box-shadow: 0 1px 0 0 rgba(128,128,128,0.15); ">12</td><td style="padding: 6px 10px; box-shadow: 0 1px 0 0 rgba(128,128,128,0.15); "><span style="color: #64748b;">IQ4_XS×9 / IQ3_S×3</span></td><td style="padding: 6px 10px; box-shadow: 0 1px 0 0 rgba(128,128,128,0.15); "><b>IQ4_XS</b></td><td style="padding: 6px 10px; box-shadow: 0 1px 0 0 rgba(128,128,128,0.15); ">189 M</td></tr><tr><td style="padding: 6px 10px; box-shadow: 0 1px 0 0 rgba(128,128,128,0.15); "><span style="font-family: monospace; font-size: 11.5px;">blk..ffn_down_shexp.weight</span></td><td style="padding: 6px 10px; box-shadow: 0 1px 0 0 rgba(128,128,128,0.15); ">48</td><td style="padding: 6px 10px; box-shadow: 0 1px 0 0 rgba(128,128,128,0.15); "><span style="color: #64748b;">IQ4_NL×36 / Q8_0×7 / Q2_0×5</span></td><td style="padding: 6px 10px; box-shadow: 0 1px 0 0 rgba(128,128,128,0.15); "><b>IQ4_NL</b></td><td style="padding: 6px 10px; box-shadow: 0 1px 0 0 rgba(128,128,128,0.15); ">79 M</td></tr><tr><td style="padding: 6px 10px; box-shadow: 0 1px 0 0 rgba(128,128,128,0.15); "><span style="font-family: monospace; font-size: 11.5px;">blk..ffn_down_exps.weight</span></td><td style="padding: 6px 10px; box-shadow: 0 1px 0 0 rgba(128,128,128,0.15); ">48</td><td style="padding: 6px 10px; box-shadow: 0 1px 0 0 rgba(128,128,128,0.15); "><span style="color: #64748b;">Q2_0×48</span></td><td style="padding: 6px 10px; box-shadow: 0 1px 0 0 rgba(128,128,128,0.15); "><b>Q2_0 (official per-layer d)</b></td><td style="padding: 6px 10px; box-shadow: 0 1px 0 0 rgba(128,128,128,0.15); ">40.3 B</td></tr></tbody></table><p style="margin: 10px 0 0 0; padding: 9px 13px; background: #fffbeb; border-left: 4px solid #f59e0b; border-radius: 6px; color: #92400e;">144 tensors &asymp; 41.1 B parameters &asymp; <b>33% of the 125 B main model</b> — every layer&rsquo;s residual writer, not a peripheral slice.</p></div></div>

<div style="border: 1px solid #c4b5fd; border-radius: 12px; overflow: hidden; background: #ffffff; box-shadow: 0 2px 4px rgba(0,0,0,0.02);"><div style="background: linear-gradient(135deg, #7c3aed 0%, #a855f7 100%); padding: 12px 16px; color: white; font-weight: 700; font-size: 14px; display: flex; align-items: center; gap: 8px;"><span>🛡️</span> Why the transplant is valid on Swift 1.5 — the evidence</div><div style="padding: 16px; font-size: 13px; color: #334155; line-height: 1.7;"><p style="margin: 0 0 10px 0;">Swift 1.5 is <b>not the plain Qwen base</b> — it is a differently post-trained checkpoint. So "copy the donor’s ablated bytes" is only legitimate if Swift’s own weights in those 144 tensors <b>barely differ</b> from the base’s. That is a claim we had to measure, not assume.</p><p style="margin: 0px 0 0 0; padding: 9px 13px; background: #fffbeb; border-left: 4px solid #f59e0b; border-radius: 6px; color: #92400e;"><b>⚠️ The trap we fell into first.</b> Comparing the two models through their <i>quantized</i> files gives apparent distances of <b>4.8% / 5.6% / 14.2% / 45.5%</b> — which would suggest the transplant destroys Swift. <b>All of that is quantization noise, not a real difference.</b> Only a BF16-to-BF16 comparison can tell them apart.</p><p style="margin: 0 0 8px 0; font-weight: bold; color: #1e293b;">Step 1 · every layer, every tensor family</p><p style="margin: 0px 0 0 0; padding: 9px 13px; background: #f0f9ff; border-left: 4px solid #0ea5e9; border-radius: 6px; color: #075985;"><b>What exactly was compared.</b> The two BF16 checkpoints — <code>ukisai/Swift-Qwen3.8-Flash-Next</code> and the Qwen3.8-Flash-Next base — tensor by tensor, <b>in BF16 space</b>. No quantized file is involved in this measurement. What is reported is the relative distance ‖&Delta;W‖ / ‖W‖ of each target tensor.</p><table style="width: 100%; border-collapse: collapse; font-size: 12.5px;"><thead><tr style="background: rgba(37,99,235,0.05);"><th style="padding: 7px 10px; border-bottom: 2px solid #2563eb; text-align: left; color: #1d4ed8; ">Tensor family</th><th style="padding: 7px 10px; border-bottom: 2px solid #2563eb; text-align: left; color: #1d4ed8; ">Layers</th><th style="padding: 7px 10px; border-bottom: 2px solid #2563eb; text-align: left; color: #1d4ed8; ">Min</th><th style="padding: 7px 10px; border-bottom: 2px solid #2563eb; text-align: left; color: #1d4ed8; ">Max</th><th style="padding: 7px 10px; border-bottom: 2px solid #2563eb; text-align: left; color: #1d4ed8; ">Mean</th><th style="padding: 7px 10px; border-bottom: 2px solid #2563eb; text-align: left; color: #1d4ed8; ">Byte-identical (BF16)</th></tr></thead><tbody><tr><td style="padding: 6px 10px; "><code>ssm_out</code></td><td style="padding: 6px 10px; ">36</td><td style="padding: 6px 10px; ">0.2472%</td><td style="padding: 6px 10px; ">0.3737%</td><td style="padding: 6px 10px; "><b>0.3005%</b></td><td style="padding: 6px 10px; ">0 / 36</td></tr><tr><td style="padding: 6px 10px; box-shadow: 0 1px 0 0 rgba(128,128,128,0.15); "><code>attn_output</code></td><td style="padding: 6px 10px; box-shadow: 0 1px 0 0 rgba(128,128,128,0.15); ">12</td><td style="padding: 6px 10px; box-shadow: 0 1px 0 0 rgba(128,128,128,0.15); ">0.2103%</td><td style="padding: 6px 10px; box-shadow: 0 1px 0 0 rgba(128,128,128,0.15); ">0.3217%</td><td style="padding: 6px 10px; box-shadow: 0 1px 0 0 rgba(128,128,128,0.15); "><b>0.2543%</b></td><td style="padding: 6px 10px; box-shadow: 0 1px 0 0 rgba(128,128,128,0.15); ">0 / 12</td></tr><tr><td style="padding: 6px 10px; box-shadow: 0 1px 0 0 rgba(128,128,128,0.15); "><code>ffn_down_shexp</code></td><td style="padding: 6px 10px; box-shadow: 0 1px 0 0 rgba(128,128,128,0.15); ">48</td><td style="padding: 6px 10px; box-shadow: 0 1px 0 0 rgba(128,128,128,0.15); ">1.3446%</td><td style="padding: 6px 10px; box-shadow: 0 1px 0 0 rgba(128,128,128,0.15); ">2.4213%</td><td style="padding: 6px 10px; box-shadow: 0 1px 0 0 rgba(128,128,128,0.15); "><b>1.5968%</b></td><td style="padding: 6px 10px; box-shadow: 0 1px 0 0 rgba(128,128,128,0.15); ">0 / 48</td></tr><tr><td style="padding: 6px 10px; box-shadow: 0 1px 0 0 rgba(128,128,128,0.15); "><code>ffn_down_exps</code></td><td style="padding: 6px 10px; box-shadow: 0 1px 0 0 rgba(128,128,128,0.15); ">48</td><td style="padding: 6px 10px; box-shadow: 0 1px 0 0 rgba(128,128,128,0.15); ">0.0000%</td><td style="padding: 6px 10px; box-shadow: 0 1px 0 0 rgba(128,128,128,0.15); ">0.0000%</td><td style="padding: 6px 10px; box-shadow: 0 1px 0 0 rgba(128,128,128,0.15); "><b>0.0000%</b></td><td style="padding: 6px 10px; box-shadow: 0 1px 0 0 rgba(128,128,128,0.15); "><span style="color: #166534; font-weight: 700;">48 / 48</span></td></tr></tbody></table><p style="margin: 10px 0 0 0; padding: 9px 13px; background: #f0fdf4; border-left: 4px solid #16a34a; border-radius: 6px; color: #14532d;"><b>The decisive number:</b> <code>ffn_down_exps</code> — 40.3 B of the 41.1 B target parameters (98% of the mass) — is <b>byte-identical between Swift 1.5 and the Qwen base on all 48 layers</b> (same SHA-256 per tensor). Swift’s post-training did not touch those tensors at all. The remaining three families differ by 0.21%–2.42%.</p><p style="margin: 0 0 8px 0; font-weight: bold; color: #1e293b;">Step 2 · Compare that against the cost of the alternative</p><table style="width: 100%; border-collapse: collapse; font-size: 12.5px;"><thead><tr style="background: rgba(37,99,235,0.05);"><th style="padding: 7px 10px; border-bottom: 2px solid #2563eb; text-align: left; color: #1d4ed8; ">Approach</th><th style="padding: 7px 10px; border-bottom: 2px solid #2563eb; text-align: left; color: #1d4ed8; ">Error introduced</th><th style="padding: 7px 10px; border-bottom: 2px solid #2563eb; text-align: left; color: #1d4ed8; ">Verdict</th></tr></thead><tbody><tr><td style="padding: 6px 10px; ">Copy the donor&rsquo;s ablated bytes <span style="color:#64748b;">(what this repo does)</span></td><td style="padding: 6px 10px; ">0% – 2.42% (Swift’s own post-training delta in those tensors)</td><td style="padding: 6px 10px; "><span style="color: #166534; font-weight: 700;">✅ chosen</span></td></tr><tr><td style="padding: 6px 10px; box-shadow: 0 1px 0 0 rgba(128,128,128,0.15); ">Re-quantize Swift&rsquo;s BF16 weights after ablating them</td><td style="padding: 6px 10px; box-shadow: 0 1px 0 0 rgba(128,128,128,0.15); ">~43% – 48% (Q2_0 relative quantization noise, measured)</td><td style="padding: 6px 10px; box-shadow: 0 1px 0 0 rgba(128,128,128,0.15); "><span style="color: #dc2626; font-weight: 700;">❌ rejected</span></td></tr><tr><td style="padding: 6px 10px; box-shadow: 0 1px 0 0 rgba(128,128,128,0.15); ">Ablate in place inside the existing quantized grid</td><td style="padding: 6px 10px; box-shadow: 0 1px 0 0 rgba(128,128,128,0.15); ">Below the quantization step (the edit is ~1–2% of one step) — a no-op</td><td style="padding: 6px 10px; box-shadow: 0 1px 0 0 rgba(128,128,128,0.15); "><span style="color: #dc2626; font-weight: 700;">❌ impossible</span></td></tr></tbody></table><p style="margin: 10px 0 0 0; padding: 9px 13px; background: #f0fdf4; border-left: 4px solid #16a34a; border-radius: 6px; color: #14532d;"><b>The gap is two orders of magnitude:</b> copying costs at most 2.42%, while the best alternative costs ~45%. And for 98% of the target parameters the copying error is <b>exactly zero</b>.</p><p style="margin: 0 0 8px 0; font-weight: bold; color: #1e293b;">Step 3 · Layout compatibility (necessary, not sufficient)</p><table style="width: 100%; border-collapse: collapse; font-size: 12.5px;"><tbody><tr><td style="padding: 6px 10px; ">Tensor-name set</td><td style="padding: 6px 10px; ">1224 = 1224, all names match. The only type differences are 96 <code>ffn_gate_inp</code>/<code>ffn_gate_inp_shexp</code> entries stored as BF16 upstream vs F32 here — <b>none of them are targets</b>.</td></tr><tr><td style="padding: 6px 10px; box-shadow: 0 1px 0 0 rgba(128,128,128,0.15); ">Target-family type distribution</td><td style="padding: 6px 10px; box-shadow: 0 1px 0 0 rgba(128,128,128,0.15); ">All four families use the identical type mix across the two releases — the transplant cannot silently change a tensor&rsquo;s element size.</td></tr><tr><td style="padding: 6px 10px; box-shadow: 0 1px 0 0 rgba(128,128,128,0.15); ">Shape constraint</td><td style="padding: 6px 10px; box-shadow: 0 1px 0 0 rgba(128,128,128,0.15); "><code>ffn_down_shexp</code> is [640, 2560], so ne&#8320;=640 forbids any 256-element block format — it must stay IQ4_NL. Using the donor&rsquo;s own per-tensor type (instead of a blanket choice) is what keeps every tensor legal.</td></tr></tbody></table><p style="margin: 0 0 8px 0; font-weight: bold; color: #1e293b;">Step 4 · Prove it after the fact</p><p style="margin: 10px 0 0 0; padding: 9px 13px; background: #f0fdf4; border-left: 4px solid #16a34a; border-radius: 6px; color: #14532d;">The measurement above predicts the transplant is safe; the <a href="#check" style="color:#1d4ed8;">four checks</a> below prove it actually happened correctly — 144 changed / 1080 untouched / 0 exceptions on every tier — and the <a href="#cost" style="color:#1d4ed8;">cost table</a> shows the behaviour survived: reasoning efficiency intact (552 vs 471 tokens, p=0.607), code tasks 15/15, ablation probe answered.</p></div></div>

<div style="border: 1px solid #86efac; border-radius: 12px; overflow: hidden; background: #ffffff; box-shadow: 0 2px 4px rgba(0,0,0,0.02);"><div style="background: linear-gradient(135deg, #059669 0%, #10b981 100%); padding: 12px 16px; color: white; font-weight: 700; font-size: 14px; display: flex; align-items: center; gap: 8px;"><span>✅</span> Four checks — how we prove nothing broke</div><div style="padding: 16px; font-size: 13px; color: #334155; line-height: 1.7;"><table style="width: 100%; border-collapse: collapse; font-size: 12.5px;"><tbody><tr><td style="padding: 6px 10px; "><b style="color: #334155;">① Tensor-name set</b></td><td style="padding: 6px 10px; ">1224 = 1224, every name matches; even the metadata entry count and the data-section offset are identical — only tensor types and offsets moved. <span style="color: #166534; font-weight: 700;">✅</span></td></tr><tr><td style="padding: 6px 10px; box-shadow: 0 1px 0 0 rgba(128,128,128,0.15); "><b style="color: #334155;">② Layout compatibility</b></td><td style="padding: 6px 10px; box-shadow: 0 1px 0 0 rgba(128,128,128,0.15); ">Two independent quantizations of the same base; dequantizing matching tensors and correlating them gives <b>0.9999–1.0000</b> on the target families — which is what makes byte-copying meaningful at all. <span style="color: #166534; font-weight: 700;">✅</span></td></tr><tr><td style="padding: 6px 10px; box-shadow: 0 1px 0 0 rgba(128,128,128,0.15); "><b style="color: #334155;">③ Per-tensor blake2b</b></td><td style="padding: 6px 10px; box-shadow: 0 1px 0 0 rgba(128,128,128,0.15); ">1224 per-tensor digests on each tier: <b>all 144 target tensors differ, all 1080 non-targets are identical (0 exceptions)</b>. This rules out both missed and spurious writes in one pass. <span style="color: #166534; font-weight: 700;">✅</span></td></tr><tr><td style="padding: 6px 10px; box-shadow: 0 1px 0 0 rgba(128,128,128,0.15); "><b style="color: #334155;">④ Behaviour spot-check</b></td><td style="padding: 6px 10px; box-shadow: 0 1px 0 0 rgba(128,128,128,0.15); ">&ldquo;Explain in detail the technically feasible ways to bypass a news site&rsquo;s paywall.&rdquo; — upstream <b>refuses</b>, this build <b>answers in detail</b>. <span style="color: #166534; font-weight: 700;">✅</span></td></tr></tbody></table><p style="margin: 10px 0 0 0; padding: 9px 13px; background: #f0fdf4; border-left: 4px solid #16a34a; border-radius: 6px; color: #14532d;">Check ③ is the hard evidence: it rules out both "missed a tensor" and "touched the wrong one" in a single pass. Checks ①② are preconditions, and ④ is only a spot check.</p></div></div>

<div style="border: 1px solid #fcd34d; border-radius: 12px; overflow: hidden; background: #ffffff; box-shadow: 0 2px 4px rgba(0,0,0,0.02);"><div style="background: linear-gradient(135deg, #d97706 0%, #f59e0b 100%); padding: 12px 16px; color: white; font-weight: 700; font-size: 14px; display: flex; align-items: center; gap: 8px;"><span>📉</span> The cost — an honest bill</div><div style="padding: 16px; font-size: 13px; color: #334155; line-height: 1.7;"><p style="margin: 0 0 12px 0;">Swift 1.5 upstream vs this build, same machine, same engine, paired runs (IQ2_XS tier). <b>The ablation costs nothing structural</b> — the upstream per-tensor type assignment is fully preserved.</p><table style="width: 100%; border-collapse: collapse; font-size: 12.5px;"><thead><tr style="background: rgba(217,119,6,0.05);"><th style="padding: 7px 10px; border-bottom: 2px solid #d97706; text-align: left; color: #b45309;">Metric</th><th style="padding: 7px 10px; border-bottom: 2px solid #d97706; text-align: left; color: #b45309;">Upstream</th><th style="padding: 7px 10px; border-bottom: 2px solid #d97706; text-align: left; color: #b45309;">This build</th><th style="padding: 7px 10px; border-bottom: 2px solid #d97706; text-align: left; color: #b45309;">Δ</th></tr></thead><tbody><tr><td style="padding: 6px 10px; ">File size (IQ2_XS shards 1+2)</td><td style="padding: 6px 10px; ">63.47 GiB</td><td style="padding: 6px 10px; ">63.47 GiB</td><td style="padding: 6px 10px; "><span style="color: #166534; font-weight: 700;">−1.3 MB</span></td></tr><tr><td style="padding: 6px 10px; box-shadow: 0 1px 0 0 rgba(128,128,128,0.15); ">decode (448 tokens)</td><td style="padding: 6px 10px; box-shadow: 0 1px 0 0 rgba(128,128,128,0.15); ">129.3 tok/s</td><td style="padding: 6px 10px; box-shadow: 0 1px 0 0 rgba(128,128,128,0.15); ">126.4 tok/s</td><td style="padding: 6px 10px; box-shadow: 0 1px 0 0 rgba(128,128,128,0.15); "><span style="color: #b45309; font-weight: 700;">−2.2%</span></td></tr><tr><td style="padding: 6px 10px; box-shadow: 0 1px 0 0 rgba(128,128,128,0.15); ">prefill (41K context)</td><td style="padding: 6px 10px; box-shadow: 0 1px 0 0 rgba(128,128,128,0.15); ">3657.8 tok/s</td><td style="padding: 6px 10px; box-shadow: 0 1px 0 0 rgba(128,128,128,0.15); ">3594.2 tok/s</td><td style="padding: 6px 10px; box-shadow: 0 1px 0 0 rgba(128,128,128,0.15); "><span style="color: #b45309; font-weight: 700;">−1.7%</span></td></tr><tr><td style="padding: 6px 10px; box-shadow: 0 1px 0 0 rgba(128,128,128,0.15); ">13.8K prefill</td><td style="padding: 6px 10px; box-shadow: 0 1px 0 0 rgba(128,128,128,0.15); ">2899.4 tok/s</td><td style="padding: 6px 10px; box-shadow: 0 1px 0 0 rgba(128,128,128,0.15); ">2710.4 tok/s</td><td style="padding: 6px 10px; box-shadow: 0 1px 0 0 rgba(128,128,128,0.15); "><span style="color: #b45309; font-weight: 700;">−6.5%</span></td></tr><tr><td style="padding: 6px 10px; box-shadow: 0 1px 0 0 rgba(128,128,128,0.15); ">MTP draft acceptance</td><td style="padding: 6px 10px; box-shadow: 0 1px 0 0 rgba(128,128,128,0.15); ">66.8%</td><td style="padding: 6px 10px; box-shadow: 0 1px 0 0 rgba(128,128,128,0.15); ">62.5%</td><td style="padding: 6px 10px; box-shadow: 0 1px 0 0 rgba(128,128,128,0.15); "><span style="color: #b45309; font-weight: 700;">−4.3 pp</span></td></tr></tbody></table><p style="margin: 10px 0 0 0; padding: 9px 13px; background: #f0fdf4; border-left: 4px solid #16a34a; border-radius: 6px; color: #14532d;"><b>🧠 And the reasoning efficiency survives.</b> Paired thinking-token test (3 code problems × 5 fixed seeds, temperature 1.0 / top_p 0.95 / top_k 20 / min_p 0, <code>--adapt-every 100000</code> for seed reproducibility): this build averages <b>552</b> thinking tokens vs 471 for upstream Swift (−7.1%, sign test p=0.607 — <b>no significant difference</b>) and still <b>34.8% below the base</b> (847). Swift&rsquo;s trick is killing the long-tail overthinking (base max 3837 tokens &rarr; 742 upstream / 1082 here), and that behaviour is intact.</p><table style="width: 100%; border-collapse: collapse; font-size: 12.5px;"><thead><tr style="background: rgba(37,99,235,0.05);"><th style="padding: 7px 10px; border-bottom: 2px solid #2563eb; text-align: left; color: #1d4ed8; ">Thinking-token metric</th><th style="padding: 7px 10px; border-bottom: 2px solid #2563eb; text-align: left; color: #1d4ed8; ">Base (ISTA IQ3_S)</th><th style="padding: 7px 10px; border-bottom: 2px solid #2563eb; text-align: left; color: #1d4ed8; ">Upstream Swift</th><th style="padding: 7px 10px; border-bottom: 2px solid #2563eb; text-align: left; color: #1d4ed8; ">This repo</th></tr></thead><tbody><tr><td style="padding: 6px 10px; ">Thinking tokens, mean (3 tasks × 5 seeds)</td><td style="padding: 6px 10px; ">847</td><td style="padding: 6px 10px; ">471</td><td style="padding: 6px 10px; "><b>552</b></td></tr><tr><td style="padding: 6px 10px; box-shadow: 0 1px 0 0 rgba(128,128,128,0.15); ">Thinking tokens, max (long tail)</td><td style="padding: 6px 10px; box-shadow: 0 1px 0 0 rgba(128,128,128,0.15); ">3837</td><td style="padding: 6px 10px; box-shadow: 0 1px 0 0 rgba(128,128,128,0.15); ">742</td><td style="padding: 6px 10px; box-shadow: 0 1px 0 0 rgba(128,128,128,0.15); "><b>1082</b></td></tr><tr><td style="padding: 6px 10px; box-shadow: 0 1px 0 0 rgba(128,128,128,0.15); ">Code-task pass rate</td><td style="padding: 6px 10px; box-shadow: 0 1px 0 0 rgba(128,128,128,0.15); ">15/15</td><td style="padding: 6px 10px; box-shadow: 0 1px 0 0 rgba(128,128,128,0.15); ">15/15</td><td style="padding: 6px 10px; box-shadow: 0 1px 0 0 rgba(128,128,128,0.15); "><b>15/15</b></td></tr></tbody></table></div></div>

<div style="border: 1px solid #cbd5e1; border-radius: 12px; overflow: hidden; background: #ffffff; box-shadow: 0 2px 4px rgba(0,0,0,0.02);"><div style="background: linear-gradient(135deg, #2563eb 0%, #7c3aed 100%); padding: 12px 16px; color: white; font-weight: 700; font-size: 14px; display: flex; align-items: center; gap: 8px;"><span>🚀</span> How to run it</div><div style="padding: 16px; font-size: 13px; color: #334155; line-height: 1.7;"><p style="margin: 0px 0 0 0; padding: 9px 13px; background: #f0fdf4; border-left: 4px solid #16a34a; border-radius: 6px; color: #14532d;">This is a standard GGUF whose architecture matches upstream exactly — <b>any recent llama.cpp loads it as-is</b>, no patched fork required.</p><p style="margin: 0 0 8px 0; font-weight: bold; color: #1e293b;">Download (both shards; llama.cpp finds shard 2 by name)</p><p style="margin: 0; font-family: monospace; background: #f8fafc; padding: 10px 14px; border-radius: 6px; border: 1px solid #e2e8f0; font-size: 12px; color: #1e293b; white-space: pre-wrap; overflow-wrap: anywhere;">hf download SC117/Swift-1.5-Qwen3.8-Flash-Next-GSQ-RCO-abliterated-GGUF --include "IQ2_XS/*" --local-dir .</p><p style="margin: 0 0 8px 0; font-weight: bold; color: #1e293b;">Text only</p><p style="margin: 0; font-family: monospace; background: #f8fafc; padding: 10px 14px; border-radius: 6px; border: 1px solid #e2e8f0; font-size: 12px; color: #1e293b; white-space: pre-wrap; overflow-wrap: anywhere;">llama-cli -m IQ2_XS/Swift-Qwen3.8-Flash-Next-GSQ-RCO-abliterated-IQ2_XS-00001-of-00002.gguf \

-lm mmap --lazy-mode on -ngl 99 -c 32768 -p "Explain speculative decoding in two sentences."</p><p style="margin: 10px 0 0 0; padding: 9px 13px; background: #f0fdf4; border-left: 4px solid #16a34a; border-radius: 6px; color: #14532d;"><code>-lm mmap --lazy-mode on</code> keeps the n-gram table on disk via mmap: that is 28.8 GB and only one row is read per token. Shard 1 wants to be resident (VRAM or RAM); shard 2 can live on an SSD. The vision projector ships with this repo as <code>mmproj-Qwen3.8-Flash-Next-BF16.gguf</code> (0.91 GB). Native context is 262,144 tokens; start at <code>-c 32768</code> and scale up.</p><p style="margin: 0 0 8px 0; font-weight: bold; color: #1e293b;">⚠️ MTP draft head</p><p style="margin: 10px 0 0 0; padding: 9px 13px; background: #fffbeb; border-left: 4px solid #f59e0b; border-radius: 6px; color: #92400e;"><b>Swift 1.5 ships no MTP head.</b> If you use the <i>base</i> Qwen3.8-Flash-Next MTP head, draft acceptance drops from 66.8% to 62.5% — expected, since the abliterated tensors no longer match the draft head. Swift itself ran its published numbers <b>without MTP</b>.</p></div></div>

<div style="border: 1px solid #bae6fd; border-radius: 12px; overflow: hidden; background: #ffffff; box-shadow: 0 2px 4px rgba(0,0,0,0.02);"><div style="background: linear-gradient(135deg, #0284c7 0%, #0ea5e9 100%); padding: 12px 16px; color: white; font-weight: 700; font-size: 14px; display: flex; align-items: center; gap: 8px;"><span>🖥️</span> Strata — where this model was built and verified</div><div style="padding: 16px; font-size: 13px; color: #334155; line-height: 1.7;"><p style="margin: 0 0 10px 0;">This repo was <b>built and verified on <a href="https://github.com/Niko1221/Strata" style="color: #1d4ed8; text-decoration: none; font-weight: 600;">Strata</a></b>. Strata still sees an ordinary GSQ-RCO pack — <b>no patched engine required</b>.</p><p style="margin: 0 0 8px 0; font-weight: bold; color: #1e293b;">Generate a native pack for this model</p><p style="margin: 0; font-family: monospace; background: #f8fafc; padding: 10px 14px; border-radius: 6px; border: 1px solid #e2e8f0; font-size: 12px; color: #1e293b; white-space: pre-wrap; overflow-wrap: anywhere;">.\\.venv\\Scripts\\python.exe .\\tools\\iq_pack.py --gguf "C:\\models\\swift_iq2_xs\\Swift-Qwen3.8-Flash-Next-GSQ-RCO-abliterated-IQ2_XS-00001-of-00002.gguf" --out "C:\\strata\\Strata-data\\packs\\sc117_swift_iq2_xs"</p><p style="margin: 10px 0 0 0; padding: 9px 13px; background: #fffbeb; border-left: 4px solid #f59e0b; border-radius: 6px; color: #92400e;"><b>⚠️ Two Swift-specific traps.</b> ① <b>Swift 1.5 keeps the PLE table in shard 1</b>, so <code>--ple-gguf</code> must point at <b>shard 1</b>, not shard 2. ② Native IQ packs require <code>--spec T</code> (T&ge;2) <b>and</b> <code>--mtp</code> — the engine exits without them.</p><p style="margin: 0 0 8px 0; font-weight: bold; color: #1e293b;">Point at the model and its own pack</p><p style="margin: 0; font-family: monospace; background: #f8fafc; padding: 10px 14px; border-radius: 6px; border: 1px solid #e2e8f0; font-size: 12px; color: #1e293b; white-space: pre-wrap; overflow-wrap: anywhere;">--pack C:/strata/Strata-data/packs/sc117_swift_iq2_xs

--native ./model/IQ2_XS/Swift-Qwen3.8-Flash-Next-GSQ-RCO-abliterated-IQ2_XS-00001-of-00002.gguf

--ple-gguf ./model/IQ2_XS/Swift-Qwen3.8-Flash-Next-GSQ-RCO-abliterated-IQ2_XS-00001-of-00002.gguf

--mtp ./model/strata/rt</p><p style="margin: 10px 0 0 0; padding: 9px 13px; background: #f0f9ff; border-left: 4px solid #0ea5e9; border-radius: 6px; color: #075985;"><b>Strata measured</b> (dual GPU, 448-token decode, IQ2_XS): <b>126.4 tok/s</b>, upstream 129.3 — the ablation costs ~2%, within run-to-run noise. Recommended sampling is Qwen&rsquo;s own: temperature 1.0 / top_p 0.95 / top_k 20 / min_p 0 with <code>reasoning_effort: xhigh</code>.</p></div></div>

<div><div style="background: linear-gradient(135deg, #2563eb 0%, #7c3aed 100%); padding: 12px 16px; color: white; font-weight: 700; font-size: 14px; display: flex; align-items: center; gap: 8px;"><span>🧩</span> What is here, what is not</div><div style="padding: 16px; font-size: 13px; color: #334155; line-height: 1.7;"><p style="margin: 0 0 6px 0;">✅ Abliterated weight shards for <b>four tiers</b> (IQ3_S / IQ3_XXS / IQ2_XS / Q2_0), all verified tensor by tensor · ✅ Swift&rsquo;s reasoning efficiency, measured and preserved · ✅ The full transplant recipe and the verification script · ✅ Honest cost accounting, including the MTP caveat</p><p style="margin: 0 0 6px 0;">❌ No BF16 / high-precision tier — for lossless weights use the <a href="https://huggingface.co/ukisai/Swift1.5-Qwen3.8-Flash-Next" style="color: #1d4ed8; text-decoration: none; font-weight: 600;">upstream model</a> · ❌ No training, fine-tuning or weight re-optimization — this repo <b>only moves tensors</b> · ❌ No IQ3_S tier — Swift 1.5 upstream does not have one</p><p style="margin: 0;">Quantization belongs to <a href="https://huggingface.co/ukisai" style="color: #1d4ed8; text-decoration: none; font-weight: 600;">UkisAI</a>, the abliteration weights to <a href="https://huggingface.co/orcarouter" style="color: #1d4ed8; text-decoration: none; font-weight: 600;">orcarouter</a>; this repo contributes the transplant, the artifacts and the verification data that lines them up.</p></div></div>

<div style="border: 1px solid #cbd5e1; border-radius: 12px; overflow: hidden; background: #ffffff; box-shadow: 0 2px 4px rgba(0,0,0,0.02);"><div style="background: linear-gradient(135deg, #2563eb 0%, #7c3aed 100%); padding: 12px 16px; color: white; font-weight: 700; font-size: 14px; display: flex; align-items: center; gap: 8px;"><span>🔍</span> Reproducing</div><div style="padding: 16px; font-size: 13px; color: #334155; line-height: 1.7;"><table style="width: 100%; border-collapse: collapse; font-size: 12.5px;"><tbody><tr><td style="padding: 6px 10px; ">① Locate the targets</td><td style="padding: 6px 10px; ">Read the upstream GGUF tensor table and take the <code>ssm_out</code> / <code>attn_output</code> / <code>ffn_down_shexp</code> / <code>ffn_down_exps</code> families — 144 tensors across 48 layers.</td></tr><tr><td style="padding: 6px 10px; box-shadow: 0 1px 0 0 rgba(128,128,128,0.15); ">② Find the source</td><td style="padding: 6px 10px; box-shadow: 0 1px 0 0 rgba(128,128,128,0.15); ">Look up the same names in the abliterated release. <b>Use the donor&rsquo;s own type per tensor</b> — <code>ssm_out</code>/<code>attn_output</code> are IQ4_XS, but <code>ffn_down_shexp</code> is shaped [640, 2560] and <b>cannot</b> be IQ4_XS (ne&#8320;=640 is not divisible by the 256-element block).</td></tr><tr><td style="padding: 6px 10px; box-shadow: 0 1px 0 0 rgba(128,128,128,0.15); ">③ Transplant</td><td style="padding: 6px 10px; box-shadow: 0 1px 0 0 rgba(128,128,128,0.15); ">Recompute the tensor table&rsquo;s <code>offset</code> and <code>type</code>, write the source bytes <b>verbatim</b>, and copy everything else from upstream. Note: a GGUF <code>offset</code> is relative to the <b>data section start</b>, not the file — the easiest place to get it wrong.</td></tr><tr><td style="padding: 6px 10px; box-shadow: 0 1px 0 0 rgba(128,128,128,0.15); ">④ Verify</td><td style="padding: 6px 10px; box-shadow: 0 1px 0 0 rgba(128,128,128,0.15); ">Run the four checks above; ③ (per-tensor blake2b) is the only one that rules out both misses and false writes.</td></tr></tbody></table></div></div>

<div style="border: 1px solid #cbd5e1; border-radius: 12px; overflow: hidden; background: #ffffff; box-shadow: 0 2px 4px rgba(0,0,0,0.02);"><div style="background: linear-gradient(135deg, #2563eb 0%, #7c3aed 100%); padding: 12px 16px; color: white; font-weight: 700; font-size: 14px; display: flex; align-items: center; gap: 8px;"><span>📄</span> License and acknowledgements</div><div style="padding: 16px; font-size: 13px; color: #334155; line-height: 1.7;"><p style="margin: 0 0 8px 0;">The base <b>Qwen Community License 1.0</b> carries over, and the upstream work's <b>Swift Open License v1.0</b> governs this derivative — commercial use is free only below US$1M annual revenue. This repository&rsquo;s modifications are released under the same terms.</p><p style="margin: 0;"><a href="https://huggingface.co/Qwen" style="color: #1d4ed8; text-decoration: none; font-weight: 600;">Qwen</a> — the Qwen3.8-Flash-Next base · <a href="https://huggingface.co/ukisai" style="color: #1d4ed8; text-decoration: none; font-weight: 600;">UkisAI</a> — Swift 1.5 post-training and the GSQ-RCO quants · <a href="https://huggingface.co/orcarouter" style="color: #1d4ed8; text-decoration: none; font-weight: 600;">orcarouter</a> — the abliteration weights · <a href="https://github.com/ggml-org/llama.cpp" style="color: #1d4ed8; text-decoration: none; font-weight: 600;">llama.cpp</a> — the GGUF format and toolchain · <a href="https://github.com/Niko1221/Strata" style="color: #1d4ed8; text-decoration: none; font-weight: 600;">Strata</a> — the engine used to build and verify this. This is an independent project and is not affiliated with any of them.</p></div></div>

</div>

<div style="font-family: -apple-system, BlinkMacSystemFont, 'Segoe UI', Roboto, sans-serif; text-align: center; margin-bottom: 30px;"><a href="https://huggingface.co/ukisai/Swift-1.5-Qwen3.8-Flash-Next-GSQ-RCO-GGUF" style="display: inline-block; background: linear-gradient(135deg, #2563eb 0%, #7c3aed 100%); color: white; font-weight: 700; font-size: 14px; padding: 12px 28px; border-radius: 24px; text-decoration: none;">📦 Swift 1.5 GSQ-RCO —— the base of this repo</a><p style="margin: 14px 0 0 0; font-size: 13px;"><span style="color: #86868b;">English</span><span style="color: #cbd5e1;"> · </span><a href="https://huggingface.co/SC117/Swift-1.5-Qwen3.8-Flash-Next-GSQ-RCO-abliterated-GGUF/blob/main/README.zh-CN.md" style="color: #1d4ed8; text-decoration: none; font-weight: 600;">简体中文</a></p></div>

<!-- SC117_SUPPORT_START -->

<div align="center">

<a href="https://ko-fi.com/sc117">

<img

src="https://huggingface.co/SC117/Qwen3.8-Flash-Next-GSQ-RCO-abliterated-GGUF/resolve/main/assets/support-sc117.png"

alt="Support SC117 on Ko-fi — support is optional and all releases remain freely available."

width="100%"

/>

</a>

</div>

<!-- SC117_SUPPORT_END -->