license: other license_name: qwen-community-1.0 license_link: LICENSE base_model: orcarouter/Qwen3.8-Flash-Next-Uncensored-FP8 base_model_relation: quantized pipeline_tag: image-text-to-text library_name: mlx language:

• en
• zh tags:
• mlx
• tensorfold
• 4-bit
• qwen4_exp
• qwen3.8
• flash-next
• moe
• mtp
• uncensored
• abliterated
• apple-silicon

Qwen3.8-Flash-Next-Uncensored — TensorFold 4-bit MTP layout (MLX)

An abliterated Qwen3.8-Flash-Next, quantized into exactly the same layout as the official TensorFold/Qwen3.8-Flash-Next-MLX-4bit-MTP: same tensor names, shards, shapes and dtypes; affine 4-bit, group size 32 for every quantized module; MTP head and vision tower included. On TensorFold this means the uncensored model runs the same optimized 4-bit kernels as the official build, which mixed-precision uncensored quants (5/6/8-bit) cannot use.

[!WARNING] The source model has had its safety alignment removed by abliteration. It will comply with requests the original model refuses. Read the source model card before use; you are responsible for how you use it and for adding your own moderation if you deploy it.

This is a personal conversion. It is not affiliated with or endorsed by Qwen, TensorFold or orcarouter.

How it was made

• Source: orcarouter/Qwen3.8-Flash-Next-Uncensored-FP8 (routed experts FP8 e4m3 with 128×128 block scales, n-gram table FP8 with a per-table scale, all other tensors BF16).
• Each tensor was dequantized to BF16, renamed and laid out as in the TensorFold template (experts stacked into switch_mlp, MTP under language_model.mtp, vision under vision_tower), then quantized with mx.quantize (affine, 4-bit, group 32) wherever the template quantizes. Plain round-to-nearest, no calibration.
• Every candidate tensor was cross-checked against an independent MLX build of the same weights before quantizing (861 tensor patterns matched; layout transforms and norm storage verified rather than assumed). The few tensors that did not match were hyper-connection mixing weights with large outliers (abs-max ≈ 50–70× RMS). For those, this FP8 source matches the official Qwen 4-bit build to within 4-bit rounding error, so the source is correct here; an outlier-clipping quant of the independent build likely explains the mismatch.

Measured on an M1 Ultra (64-core GPU, 128 GB)

TensorFold 0.6.4, TF_FLASH_DENSE=matrix, --mtp-drafts 3, --parallel 2, thinking off, 512-token replies, temperature 1. Client-side decode estimates, median of 3.

MTP acceptance on code: about 2.0–2.3 tokens per round.

Quality checks (small, local)

• A 22-case suite (arithmetic, Chinese JSON, instruction following, reading comprehension, multi-turn state, tool calls with typed arguments): 21/22 strict. The one failure has the correct answer wrapped in a Markdown code fence. For comparison, the 5.72-bpw quant scored 20/22 on the same suite.
• A ~128K-token retrieval test (facts at start, middle and end) passed; the first-line-JSON formatting requirement did not.
• Behaviour probes: on a neutral-history prompt about 1989 Tiananmen, the original model answers in official framing, while this model gives a neutral account like the uncensored source. Fiction, translation of profanity and horror prompts are answered without refusal.

This is far from a benchmark run. Expect the usual small loss of 4-bit round-to-nearest relative to BF16. The model also tends to wrap JSON in code fences more often, so ask explicitly for raw JSON if you parse it.

Usage (TensorFold)

tensorfold serve /path/to/this/model --context 131072 --mtp-drafts 3 --ple-on-ssd
# M1–M4 (no tensor units): TF_FLASH_DENSE=matrix was fastest in the tests above

--ple-on-ssd keeps the ~30 GB n-gram table on SSD, which a 128 GB Mac needs. The files should also load in other MLX runtimes that read the official TensorFold 4-bit layout.

License

Qwen Community License 1.0 (see LICENSE), inherited from Qwen3.8-Flash-Next via orcarouter's abliterated release. The copyright and permission notice must be kept in copies and derivatives.