IHaveNoClueAndIMustPost/Ling-3.0-flash-SUPER-GGUF overview

---

base_model:

• inclusionAI/Ling-3.0-flash

---

Mixed-precision GGUF quants of

inclusionAI/Ling-3.0-flash.

Zero slop TLDR: Just two vibe-quanted versions attempting to optimize size vs performance. Only recommended as temporary replacements until the pro-quanters release their dynamic versions.

Two builds are provided that cover two VRAM budgets, 64GB with some slight offloading and 96GB+ without.

| File | Size | Strategy | PPL vs F16 | Best for |

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

| Ling-3.0-flash-f16.gguf | 243,268 MiB (~237.6 GiB) | Full precision (baseline) | 4.3532 (reference) | Full size. N/A |

| Ling-3.0-flash-SUPER-Q4_K_M.gguf | 82,476 MiB (~80.5 GiB) | Q8_0 signal path + Q4_K expert bulk | 4.3849 (+0.73%) | 96-128+ GB VRAM (long context); 64 GB with heavier offload |

| Ling-3.0-flash-SUPER-Q3_K_M.gguf | 64,183 MiB (~62.7 GiB) | Q6_K signal path + Q3_K expert bulk | 4.5289 (+4.04%) | ~64 GB (2×32 GB), light offload |

| Ling-3.0-flash-SUPER-Q2_K.gguf | 61,776 MiB (~60.3 GiB) | Q8_0 signal path + Q2_K expert bulk | 4.9592 (+13.92%) | Failure reference. N/A |

Reference: the full-precision F16 model scores PPL = 4.3532 ± 0.0256 on the

wikitext-2 test set.

Performance

On the dual V100 32GB target system the diffence in TG was approx 10-15% between the different models, hovering around 30-35t/s with ngram drafting. This is an "e-waste" system with some non-standard/non-recommended build decisions so the numbers are likely worst rather than best case scenario.

Known model quirks

• Repetition looping. Ling-3.0-flash can occasionally fall into a repetition loop.

This is inherent to the full F16 model, not introduced by quantization, and is

much rarer in the latest revision of the model. Standard repetition penalties /

sampler settings apply.

• MTP / speculative-decoding speed. The model's native MTP (multi-token

prediction) path is reported to run slower than expected, which is mentioned in the PR and verified on my own system.

The launch commands below therefore use ngram

speculative decoding (--spec-type ngram-mod) only.

• Recommendation: try the MTP and ngram combination first with ``--spec-type ngram-mod,draft-mtp``, then ngram-only, and compare the results.

---

Ling-3.0-flash-SUPER-Q4_K_M (82,476 MiB, ~80.5 GiB)

Quantization method. Mixed-precision k-quant, no imatrix (deliberate — a uniform,

calibration-free quant avoids biasing a general-purpose model toward one corpus). The

"SUPER" tiering keeps the precision-sensitive signal path high and spends the savings on

the expert bulk:

• Q8_0 — the signal path: all attention (q/k/v/output/gate + MQA k_b/v_b/

kv_a_mqa), the gated-linear-attention (SSM) projections, the shared expert, the

embedding, the MTP head, and the input/output edge experts (blocks 2-3 and 39-42).

• Q6_K — a periodic set of down ("cycle-down") experts spaced through the stack,

plus the two dense-block down projections; these act as high-fidelity checkpoints that

stop quantization error accumulating down the model.

• Q4_K — the remaining (middle) routed-expert bulk, where most of the size is saved.
• F32 — all norms, routers, and small SSM params (auto-preserved); F16 — the

final output.weight.

Type distribution (938 tensors): 436 Q8_0, 96 Q4_K, 15 Q6_K, 390 F32, 1 F16.

Perplexity (wikitext-2 test set, n_ctx=512):

• PPL = 4.3849 ± 0.0258 vs F16 4.3532 → +0.0317 abs / +0.73%.
• Full per-chunk logs (in this repo): Ling-PPL-Q4_K_M.txt (baseline Ling-PPL-F16.txt).

Launch (2×32 GB V100 / 64 GB total, reference config):

./llama-server -fa 1 -m Ling-3.0-flash-SUPER-Q4_K_M.gguf \
  --host 0.0.0.0 --port 8080 -c 131072 -ngl 99 \
  --load-mode none --parallel 1 -ctxcp 96 -cram -1 \
  -ub 1024 -b 4096 --spec-type ngram-mod \
  -ctk q8_0 -ctv q8_0 -t 8 -ts 7,4 -ncmoe 14

---

Ling-3.0-flash-SUPER-Q3_K_M (64,183 MiB, ~62.7 GiB)

Quantization method. Same tensor roles as the Q4_K_M, but the whole precision curve

is flattened one step: the expensive top is demoted in order to pay for a

higher-precision bulk, which is where the quality actually lives.

• Q6_K — the signal path + edge experts (same set that is Q8_0 above).
• Q5_K — the "cycle-down" experts + dense down projections.
• Q3_K — the middle routed-expert bulk.
• F32 — norms / routers / SSM params; F16 — output.weight.

Eight attn_k_b tensors keep Q8_0: their shape cannot be represented by any K-quant,

so they fall back to the nearest compatible block type.

Type distribution (938 tensors): **428 Q6_K, 92 Q3_K, 15 Q5_K, 8 Q8_0, 4 Q4_K, 390 F32,

1 F16**.

Perplexity (wikitext-2 test set, n_ctx=512):

• PPL = 4.5289 ± 0.0267 vs F16 4.3532 → +0.1757 abs / +4.04%.
• Full per-chunk logs (in this repo): Ling-PPL-Q3_K_M.txt (baseline Ling-PPL-F16.txt).

Launch (2×32 GB V100 / 64 GB total, reference config):

./llama-server -fa 1 -m Ling-3.0-flash-SUPER-Q3_K_M.gguf \
  --host 0.0.0.0 --port 8080 -c 131072 -ngl 99 \
  --load-mode none --parallel 1 -ctxcp 96 -cram -1 \
  -ub 1024 -b 4096 --spec-type ngram-mod \
  -ctk q8_0 -ctv q8_0 -t 8 -ts 7,6 -ncmoe 5

Credits

For great models and quantization guidance:

<br>ubergarm

<br>AesSedai

<br>bloomer010

<br>AtomicChat

For actually doing the heavy lifting:

<br>Qwen-chan

For just tossing some prompts at an LLM:

<br>Me. I did that.