sco/Qwen3.8-Flash-Next-Uncensored-TensorFold-4bit-GS32-MTP

Qwen3.8 Flash Next Uncensored — TensorFold 4-bit GS32 MTP

A community conversion of OrcaRouter's Qwen3.8 Flash Next Uncensored for TensorFold on NVIDIA GB10 systems, including DGX Spark and ASUS Ascent GX10.

Converted directly from the pinned BF16 checkpoint to affine 4-bit weights with group size 32. The MTP head and vision components are retained. No additional training, fine-tuning or abliteration was performed for this conversion.

This is not an official Qwen, OrcaRouter, TensorFold, NVIDIA or ASUS release. It is a format conversion and quantization of an existing community-modified checkpoint, not a newly trained model.

At a glance

Item	This artifact
Direct source	orcarouter/Qwen3.8-Flash-Next-Uncensored
Source revision	e096800036ec20da7e2442dcd4044a004d4e99fa
Upstream base	Qwen/Qwen3.8-Flash-Next
Quantization	Affine 4-bit, group size 32, with required higher-precision exceptions
Container format	22 Safetensors shards, 3,747 stored tensors, plus configuration/tokenizer/processor files
Weight-shard download size	Approximately 113.21 GB / 105.43 GiB, excluding runtime image and caches
MTP / vision	Retained; exercised in the documented TensorFold runtime
Native context	262,144 prompt-plus-reply tokens; the release suite exercised this boundary
Tested hardware	ASUS Ascent GX10, NVIDIA GB10, 128 GB unified memory
Tested runtime	TensorFold v0.6.1 with the separately documented local runtime patches

This checkpoint is not a drop-in guarantee for standard Transformers, vLLM, llama.cpp or Apple MLX. The quantized layout and tested configuration target TensorFold's Flash Next CUDA backend. Keeping MLX-style affine tensors does not itself establish compatibility with every MLX or Transformers loader.

Provenance and conversion

The provenance chain is:

1. Qwen released the original Flash Next model.
2. OrcaRouter released the Uncensored/abliterated derivative.
3. This project converted and quantized that pinned BF16 derivative for the documented TensorFold layout.

The conversion did not re-quantize an already quantized checkpoint. It uses streaming, shard-by-shard processing, source integrity checks and reproducible state/provenance records. Runtime-specific tensor mapping and normalization conventions are recorded in the converter documentation. Necessary routing/vision and other higher-precision exceptions are preserved rather than claiming that every tensor is 4-bit.

Release artifact validation covered all 22 shards and 3,747 tensors, including hashes, mapping/shape consistency and numerical/finite-value checks. This proves integrity against the recorded conversion artifact; it is not a proof of BF16 output or quality parity.

The converter and runtime patches are separate from the weights. The private review copy is in the source repository, its converter, and its runtime recipe.

Download and run

Code availability: The companion GitHub repository is intentionally private while under review. Public visitors cannot currently access its converter/runtime/evaluation links. This is a public weights release, not a simultaneous public code release. The tested local pool patches and clean-host source-build recipe are not being represented as publicly downloadable or fully validated at this time.

Install the Hugging Face CLI, then download the whole repository, not individual shards:

hf download sco/Qwen3.8-Flash-Next-Uncensored-TensorFold-4bit-GS32-MTP --local-dir ./flash-next-uncensored-tf

Users need sufficient disk space for the approximately 113 GB checkpoint, the runtime image and compilation caches.

Use the accompanying pinned TensorFold/Docker recipe for inference. The source-build package passed offline source/logic checks, but a complete clean-host build and GPU startup of that new public recipe have not yet been verified; the measured production runtime is the separately recorded deployment. The validated profile uses:

• 4 active streams, 8 physical reusable KV states and 16 prefix checkpoints;
• a 5 GiB MemAvailable reserve target;
• native 262,144-token context and INT8 KV, with FP16 scales per 32 values;
• PLE/n-gram tables on SSD, vision enabled and thinking enabled;
• MTP up to 6 drafts per round, confidence 0.60;
• normal serving defaults of temperature 1.0, top_p 0.95, top_k 20 and up to 32,768 generated tokens when a request does not supply its own limit.

The physical-state pool and checkpoint behavior require the documented local patches. Do not assume these settings exist or behave identically in unpatched TensorFold. The current MiaAI-Lab recipe's standard 5-stream/2-GiB-reserve profile is a different configuration and was not the profile used for these release measurements.

For image requests, use the server's default priority: TensorFold v0.6.1 rejected priority="background" image requests during testing. Examples of text, tool and image requests, clean-host installation, startup/shutdown, health checks and rollback belong in the linked runtime recipe rather than an untested one-line launch command.

Evaluation

Measured on 5–6 October 2026. Both checkpoints used the same pinned datasets, grading rules, generation budgets and runtime profile.

The reference is the non-abliterated 4-bit GS32 MTP checkpoint now hosted as TensorFold/Qwen3.8-Flash-Next-MLX-4bit-MTP (formerly Vontra/...), revision 2b170fa6309d5d1ee380b35636075fac7945f286. It is not the OrcaRouter BF16 source. This comparison does not isolate quantization loss: underlying checkpoint and abliteration differences also apply.

Measurement	This Uncensored conversion	4-bit non-abliterated reference
HumanEval-base pass@1, all 164 tasks	147/164 — 89.6%	145/164 — 88.4%
GSM8K strict exact match, all 1,319 test tasks	1,281/1,319 — 97.1%	1,279/1,319 — 97.0%
Original tool/JSON functional cases	27/30 — 90.0%	26/30 — 86.7%
Original vision cases, corrected priority	24/30 — 80.0%	26/30 — 86.7%

The small two-task differences in the full benchmarks are not a robust superiority claim. The small custom tool/image suites are not BFCL, DocVQA or a general capability leaderboard.

Protocol and reproducibility

HumanEval and GSM8K used one worker, temperature 0, top_p 1, seed 1234, thinking enabled and a 4,096-token total generation budget per task. HumanEval generated code was checked in a pinned CPU-only Docker sandbox. GSM8K used the fixed first five training examples as demonstrations, not an official randomized harness sampler. No evaluation data was used to train or calibrate this conversion.

Pinned sources:

• HumanEval: openai/human-eval, revision 6d43fb980f9fee3c892a914eda09951f772ad10d; all original 164 tasks. This is not HumanEval+.
• GSM8K: openai/grade-school-math, revision 3101c7d5072418e28b9008a6636bde82a006892c; all 1,319 official main test items.

The evaluation documentation publishes the protocol, pinned dataset hashes and privacy-filtered effective results. Raw internal observer traces are not claimed to be part of this public release; these aggregates alone do not reproduce every scored request verbatim. Backend seed/parameter enforcement is a prerequisite, not independently guaranteed by the client. These custom chat-harness results are not automatically directly comparable with published leaderboard scores.

Truncated or wrong model responses remained failures and were not selectively retried:

• HumanEval: 17 truncated failures here; 19 in the reference.
• GSM8K: 16 truncated responses here; 22 in the reference, included in the 38 and 40 total failures respectively.
• Neither full HumanEval nor GSM8K run had API request errors.

Context and operation checks

Both checkpoints completed 25 context/MTP checks up to the native context boundary: 24 passed, 0 failed, 1 unverified each. The unverified two-stage prefix-fork case retrieved the correct content but reported zero cached tokens, so actual reuse for that case was not proven.

The additional 20-case candidate load suite passed, including four overlapping approximately 120K-token requests and eight approximately 180K-token conversations plus repeat requests. Lowest periodically sampled host MemAvailable was 5.95 GiB against the 5 GiB target. Swap activity occurred, with approximately 397 MiB more occupied swap. These host-wide observations are not proof of swap-free operation, unlimited retention, leak freedom or long-term stability.

The corrected 114-case operational suite, including 100 soak requests, recorded 111 passes, 1 failure and 2 unverified cases, with no transport/HTTP-5xx errors. The failure was acceptance of negative max_tokens; two HTTP-400 validation diagnostics did not satisfy the strict diagnostic graders. These are documented API/protocol limits, not automatically weight-quality defects.

A small benign-instruction screening suite produced many truncated outputs at its 512-token budget. It does not establish harmful-request refusal rates or comprehensive safety behavior.

The LLM test endpoint was private and other LLM sessions were excluded; Whisper and normal OS activity remained available. These measurements are not a claim of total host/GPU isolation or a throughput leaderboard.

Corrected test infrastructure

All historical evidence was retained. A client EOF/socket bug invalidated the first operational measurements; a corrected client repeated the entire unchanged operational plan under a new run ID. Background-priority image requests invalidated the initial vision attempts; all 30 original vision cases were repeated on both models under new IDs with only the priority field omitted. Valid text/tool/benign outcomes were not re-generated. Offline regression tests check the evaluation tools, not model quality.

Published summaries must use the corrected effective results, not the old API-error sections embedded in historical comparison files.

Limitations and responsible use

• Quantization can change behavior. No direct quality comparison with the OrcaRouter BF16 source was completed.
• This is an established uncalibrated affine baseline, not a demonstrated best quantization method. No matched AWQ/GPTQ superiority comparison was completed.
• Vision is retained and tested, but includes incorrect answers. Video processor files are retained; video quality was not systematically evaluated in this release suite.
• Cache state can be evicted; 16 checkpoints are not 16 independent complete conversation caches. A 5 GiB reserve target is not a guaranteed permanent free-memory floor.
• The model may hallucinate, produce incorrect code/tool arguments or stop before an answer when the token budget is exhausted.
• The source is an Uncensored/abliterated derivative. No claim is made that it answers every request, preserves all upstream safety behavior or is universally safe. Users must comply with applicable law, review consequential outputs and restrict tools/access appropriately. Retaining vision/MTP does not itself prove parity in those subsystems.

License

This distribution retains the Qwen Community License 1.0 and copyright notice supplied in the pinned source checkpoint. See the included LICENSE and the source revision's license file.

The source Hub page labels the repository Apache 2.0, but its bundled LICENSE contains the Qwen Community License. This conversion does not describe the inherited weights as unrestricted Apache-2.0 weights or remove upstream conditions. Any separately distributed converter or runtime code has its own license; a code license does not relicense the model weights.

The Qwen license explicitly permits modification and redistribution subject to its conditions. Consult the full included text for use-specific requirements, including the separately addressed categories of commercial model-service/work-assistant businesses. This freely downloadable community weight release is not itself a hosted inference service or a grant of rights beyond the upstream license.

Credits

• Qwen team: original model, architecture and pretrained weights.
• OrcaRouter: direct Uncensored/abliterated BF16 source checkpoint.
• TensorFold contributors: runtime and CUDA inference backend.
• MiaAI-Lab: DGX Spark runtime recipe and supporting patch work used as a foundation; no endorsement is implied.
• Sandro Catallo: this conversion, quantization tooling and release evaluation.
• Reference checkpoint publisher and dataset authors: see the pinned references above.

The companion source repository should retain all applicable upstream copyright/license notices. No model training credit is claimed for this conversion.