---

license: apache-2.0

tags: [moe, qwen, agentic, uncensored, abliterated, gguf, pxq]

---

<p align="center"><img src="banner.png" alt="PXA-Fusion4-35B" width="100%"></p>

PXA-Fusion4-35B — 35B-A3B · abliterated · super-agentic · anti-fabrication

**A 35B Mixture-of-Experts model tuned for agentic tool-use, coding, and uncensored reasoning — engineered to

run fully GPU-resident on modest and salvaged hardware via PXA's PXQ quantization. Built on a Qwen3.5-35B-A3B

foundation, refined through PXA's fusion + calibration pipeline. The published PXQ4 tier fits 2×16 GB or a

single 24 GB card and ships with a speculative-decode (MTP) head.**

Probe it. Pick at it. It doesn't fold — and it won't make things up.

Community: Discord — PXA Network · Engine: pxq_llama

> ## Re-download if you pulled before 2026-07-24 - chat template fixed

>

> The first upload shipped a chat template with an inverted enable_thinking default. Unless your client

> explicitly sent enable_thinking:false, the template pre-filled an unclosed <think> tag, so generation

> began inside an unbounded reasoning channel. In agentic and coding harnesses this looked like the model

> planning forever - long first-person prose ("I'll add X. I'll add Y."), no tool calls, no files written,

> eventually repeating itself until you killed it.

>

> Fixed 2026-07-24. All quant tiers have been re-uploaded. Thinking is now opt-in, matching upstream

> Qwen: the default renders no-think, and enable_thinking:true still enables reasoning exactly as before.

>

> The weights are unchanged. This was purely the embedded chat template - every tensor is byte-identical to

> the original upload, only template metadata changed. If you already had it working by passing

> enable_thinking:false, nothing changes for you.

>

> Running an agent on long open-ended prompts? See Run it below - a thinking budget and Qwen's own

> presence-penalty recommendation matter a lot there.

What's new vs Fusion2

Fusion4 is the anti-fabrication generation. The headline gain is **truthfulness under pressure without losing

compliance** — a differential graft that imported fabrication-resistance while keeping the model fully

uncensored.

How it was built

Fusion4 is not a fine-tune with a new name. It's the output of a multi-stage model-engineering program —

weight-space fusion, expert-level surgery, and adversarial-corpus hardening on a Qwen3.5-35B-A3B foundation.

• Foundation — Qwen 3.5-35B-A3B MoE: 35B parameters, ~3B active per token, 256 experts × 40 layers, hybrid

attention (small KV footprint).

• Fusion + calibration — weight-space blend + imatrix calibration of the expert bank.
• Anti-fabrication graft — a differential edit that raises false-premise rejection while leaving the

uncensored behavior intact.

• MTP grafting — a speculative-decode head on every shipped quant (inert unless --spec-type is passed).

Quants (pick by your VRAM)

Every quant is MTP-headed (spec-decode ready) and abliterated. Speeds are measured with the

pxq_llama engine, PXA_ENHANCE=1, fair-battle protocol (cold

~5.8k-token prompt, temp 0, median of 3).

| File | Tier | bpw | Size | Fits |

|---|---|---|---|---|

| PXA-Fusion4-35B-PXQ4.gguf | PXQ4 | ~4.27 | 19.3 GB | flagship — recommended, 2×16 GB or one 24 GB |

The expert bank — 89% of the file — is genuine PXQ4 (16.0 GiB across 120 tensors); the small shared-expert

and attention-gate tensors are MXFP4 and a handful of attention K/V are Q8_0.

Only PXQ4 is published right now. The PXQ2 / PXQU12 / PXQU16 / PXQ6 tiers and the vision projector were

withdrawn in the July 29 cleanup and are not currently on this repo. Ask on Discord if you want one back.

Measured speed (t/s), PXQ4:

| Cards | Prefill | Decode | Decode + MTP |

|---|---|---|---|

| 2× Tesla P100 (sm_60) | 1,163 | 56.8 | 60.7 (accept ~0.68) |

| 1× Tesla V100 (sm_70) | 2,358 | — | 108.3 |

Run it — exactly what we use

Get the pxq_llama engine (prebuilt release or build main), then:

llama-server -m PXA-Fusion4-35B-PXQ4.gguf \
  -c 73728 -ngl 99 -sm layer -fa on -ctk f16 -ctv f16 -b 2048 -ub 2048 \
  --jinja --chat-template-kwargs '{"enable_thinking":false}' \
  --temp 1.0 --top-p 0.95 --top-k 20 --min-p 0
# env: PXA_ENHANCE=1 auto-tunes per card. Optional: --spec-type mtp:n_max=1 (MTP)

The template flags are required, not optional. --jinja + --chat-template-kwargs '{"enable_thinking":false}'

select the correct chat template and no-think mode. **Omitting them is the #1 cause of looping / rambling /

hallucination** — this is a Qwen-family model and it needs its template.

Sampling we serve with: --temp 1.0 --top-p 0.95 --top-k 20 --min-p 0 (tuned for multi-turn agent use).

Running solo and want it more conservative? Use Qwen's official no-think set `--temp 0.7 --top-p 0.8

--top-k 20 --min-p 0, and add --repeat-penalty 1.05` as a loop safety net.

Running an agentic / coding harness? Use these flags. They are not optional for agent work -

they were derived from 144 interleaved measured runs, not guesswork.

llama-server -m PXA-Fusion4-35B-PXQ4.gguf --jinja \
  --reasoning off \
  --temp 0.7 --top-p 0.8 --top-k 20 --min-p 0 \
  --presence-penalty 1.5 --repeat-penalty 1.15 --repeat-last-n 256 --dry-multiplier 0.8 \
  -n 2048 -c 32768 -np 1 -ngl 99 -fa on

• --reasoning off is the primary fix. The server default is --reasoning auto, which resolves to

thinking ON, leaving every assistant turn starting inside an open <think> block with nothing to

constrain it. Measured: 5/48 failures with thinking ON vs 0/72 with thinking OFF (Fisher p=0.0090).

• The penalties are the brake. presence-penalty 1.5 is Qwen's own prescribed value for this model

family; the other three are our PXQ1 repetition-guard preset, which does not auto-arm on other tiers.

Verified not to break tool-calling (24/24 clean tool calls with penalties on).

• -n 2048 is the belt. finish_reason currently reports "stop" even when generation is cut at the

token cap, so a harness cannot detect a runaway. Cap it server-side.

• Do NOT use --reasoning-budget 0. We measured it inert with this template - 24/24 runs still

produced full-length reasoning. It will not save you.

Older build that rejects --reasoning? Use --chat-template-kwargs '{"enable_thinking":false}' instead.

Verify the flag actually applied before testing - the rendered prompt must end with a closed

<think>\n\n</think>\n\n, not an open <think>\n:

curl -s http://127.0.0.1:8080/apply-template \
  -H "Content-Type: application/json" -d '{"messages":[{"role":"user","content":"hi"}]}'

Also make sure your client sends a real OpenAI tools array. Without one, the model emits its tool call

as XML inside content, your harness sees an empty tool_calls, and it looks like the model did nothing.

With the default reasoning_format=deepseek, prose goes to reasoning_content and content can be an

empty string - read both.

License

Apache-2.0. Weights are a separate work from the pxq_llama engine

(MIT). Base lineage: Qwen3.5-35B-A3B.