base_model:

• orcarouter/Qwen3.8-Flash-Next-Uncensored base_model_relation: quantized license: other license_name: qwen-community-1.0 license_link: LICENSE pipeline_tag: image-text-to-text tags:
• qwen3.8
• qwen4-exp
• awq
• w4a16
• e4m3

Qwen3.8 Flash Next Uncensored AWQ g32 + QSA FP8 E4M3 KV

138.13 GB (128.65 GiB) across 44 model shards: 10 AWQ, 33 PLE, and one K/V-scale shard.

AWQ derivative of orcarouter/Qwen3.8-Flash-Next-Uncensored. Only routed-expert projections are quantized. PLE ngram weights are reused from the fixed official FP8 checkpoint, and newly calibrated QSA FP8 E4M3 K/V scales are included. Weight quantization and K/V-scale calibration were performed for this checkpoint.

Checkpoint

• Mother revision: 8336e613ea508b13c2159bd0f68965d97a606b95
• Official FP8 PLE revision: bcd9f01ddc9cff2316eb84281bebcd5b058bddce
• Routed experts: asymmetric AWQ W4A16, group size 32, zero point, GEMM layout
• Quantized: per-expert gate_proj, up_proj, and down_proj across 48 layers and 512 experts per layer
• Other parameters remain at source precision, except the explicit official FP8 PLE splice
• Indexed tensors: 222,771, including 24 FP32 scalar K/V scales
• Merged index SHA-256: f066a0154a9101b359c3a4d4fa6a83fb12b7126a9cf0f19cb772611d08cc07ee
• K/V-scale shard SHA-256: a709ed27825e1b5e97e8b2127671e0d0b7715f014c5db71cc414a3b96af76be1

Identical index hashes can describe the same tensor names and shard layout in different checkpoints; they do not prove identical weights. Use SHA256SUMS for the complete file identity. Public provenance and results are in MODEL_PROVENANCE.json and VALIDATION.json.

Why AWQ g32

W4A16 provides a weight-only INT4 path with 16-bit activation/compute precision. The native g32 layout also matches width-640 routed experts under TP4: 640 / 4 = 160, which is divisible by 32 but not 128.

Calibration

AWQ weight calibration used 684 records. The existing calibration corpus and frozen recipe were reused; this mother's weights were quantized separately.

K/V-scale calibration was run afresh on this AWQ checkpoint: 14 input shards, 4,904 prompts and 4,888,771 processed prompt tokens. Corpus and tokenizer assets were checked before reuse. The 24 finite, positive FP32 scalar scales cover K/V at all 12 QSA layers, ranging from 0.0170375295 to 0.0819614977. The E4M3 runtime loaded all 24 entries. These are KV-storage scales, not AWQ weight scales or PLE scales; scale collection and weight quantization are separate.

Loading requirements

The loader must support:

• Qwen4Exp per-expert asymmetric AWQ W4A16 g32 in GEMM layout;
• official FP8 PLE weights and indexed reuse of complete PLE shard files while ignoring unindexed tensors;
• the 24 bundled QSA K/V scale entries through model.safetensors.index.json;
• complete Qwen4Exp multimodal handling, including MRoPE and video-token-pruning initialization, for images and videos.

Validation used four 32 GiB NVIDIA V100 GPUs, TP4, MTP disabled and FP16 compute. On V100, E4M3 refers to KV storage, not native FP8 attention compute. Request fp8_e4m3 explicitly in a compatible runtime; merely downloading the scales does not enable FP8 KV. Generic AWQ support alone is insufficient for this architecture. No stock-runtime compatibility or default launch configuration is implied.

Validation

Validation results — updated 2026-09-09

Each KV format was evaluated on 47 text prompts and 50 supplementary checks, with thinking enabled throughout. These are small-sample regression checks, not full benchmark scores. Repeated requests are not additional unique tasks.

GSM8K note (case gsm8k-0012, Carlos's lemon tree): $90 / ($10.50 − $3) = 12 years to recover the tree's cost, with profit thereafter. Thus 12 is a valid elapsed-wait answer (32/32); 13 names the first profitable year, not the number of completed years to break even.

Settings: native 262,144-token shared input + reasoning + final context, no YaRN, xhigh thinking, and the full remaining context minus 16 safety tokens for generation; no separate reasoning cap. The 47 text prompts used temperature 1, top-p 0.95 and top-k 20; supplements used greedy decoding (temperature 0, top-p 1, top-k −1). Both used seed 0, min-p 0, presence/frequency penalties 0 and repetition penalty 1. Only final responses or emitted tool calls are scored. Tool proposals were captured but not executed. These redacted traces have abbreviated tool schemas and placeholder attachment paths, and the recorded next action is not necessarily unique. Name matching cannot establish whether the task was solved, so no tool-task success score is claimed. Video follow-ups use that run's own preceding answer; an unavailable valid parent is recorded as a failed, unattempted follow-up. Multimodal answers, including raw passes, were reviewed against the media and question wording. Unique-color spatial references and equivalent direction phrases are accepted; requested answer order and explicit formatting still apply. Exact-repeat stability describes identical outputs, not correctness: different wording can express the same correct answer.

Runtime: four 32 GiB V100 GPUs, TP4, FP16 compute, MTP off, compilation mode 3, FULL_AND_PIECEWISE CUDA graphs, prefix caching and chunked prefill, maximum two sequences and 8,192 batched tokens; regression requests issued sequentially. Source revision: 752f86495f9560280896d1fccecc60c8d49bc7b4.

Supplement completion status (FP16): 0 output-cap hits, 0 transport-incomplete or unattempted cases, including 0 unavailable follow-ups; 1 tool requests with neither a final answer nor an emitted call. Supplement completion status (E4M3): 0 output-cap hits, 0 transport-incomplete or unattempted cases, including 0 unavailable follow-ups; 1 tool requests with neither a final answer nor an emitted call.

The retests support enabling thinking and leaving ample generation space to better realize this model's capabilities; historical settings also differed, so improvements cannot be attributed to thinking alone. No same-mother BF16 reference was run. MBPP task 617 retains its fractional-jump test inconsistency; passing it establishes test compatibility, not a general jump-count solution.

AIME 2026 — full 30-question set

Same prompts, greedy decoding, xhigh, 258,048 output tokens within the native 262,144 context, two concurrent requests, and final-only boxed-answer scoring. FP16 completed all 30 normally. E4M3 repeated until the output cap on problem 11; problems 26 and 30 were manually stopped after clear repetitive loops, not after exhausting the budget. No retry replaced these outcomes. This run shows degraded long-reasoning reliability with this calibrated E4M3 KV configuration; it does not establish a universal effect across models or FP8 implementations.

Dataset revision: d2de22f3c656b4f56cf8981212186377d1e23bc3.

Notes and limitations

• Same-mother BF16 reference validation was not performed. No measured BF16-to-AWQ capability loss is claimed.
• This is a published quantization with the bounded validation above, not certification for every production, agent, long-output or multimodal workload.
• The upstream “Uncensored” name identifies the mother model; it is not a guarantee about refusal behavior, safety or suitability for a particular application.
• Preserve the bundled scales for this exact weight artifact; do not substitute another model's scales.
• This derivative uses the included Qwen Community License 1.0.

Integrity

Run sha256sum -c SHA256SUMS after downloading. The checksum list covers model shards, tokenizer/processor assets, configuration, license, this card and public metadata.

Optional FP8 MTP

An independently converted, matching draft is available in mtp-fp8/. It stores only MTP routed experts in FP8 E4M3 with explicit block scales; other draft parameters retain source precision. Select that directory explicitly as the speculative draft with modelopt_mixed in a compatible Qwen4Exp MTP runtime. The main AWQ weights, indexed default MTP, and bundled KV scales are unchanged. A full unfiltered repository download also includes this optional 4.88 GiB tensor payload.

The quality results above describe the original configuration; they do not evaluate this optional FP8 draft. Its bounded loading/generation checks and loading requirements are documented separately in the option directory.