Qwen3.8-Flash-Next-Uncensored-oQ5e-mtp

Qwen3.8-Flash-Next-Uncensored-oQ5e-mtp

[!NOTE] Updated 2026-11-04: Re-quantized with oMLX v0.7.3. Please re-download if you have an earlier copy. This is a new revision of the same model, not a new model; see What changed in this revision.

An oMLX-native oQ5e quantization of orcarouter/Qwen3.8-Flash-Next-Uncensored, the BF16 abliterated build of Qwen3.8-Flash-Next. Targets a 128 GB Apple Silicon Mac and preserves the vision tower, the MTP head, and the checkpoint's native 262,144-token architecture setting.

For the abliteration method, refusal-direction details, and MTP-consistency notes, see the source model card linked above. This repository only documents the quantization.

Companion build: LMLiutenant/Qwen3.8-Flash-Next-Uncensored-oQ6e-mtp (higher precision, ~6.9 bpw).

Nature of this model

This is an uncensored (abliterated) model: the source weights have had their refusal direction removed, so it will attempt most requests without the safety refusals present in the official Qwen release. It has no added guardrails. You are responsible for how you use it and for complying with the license and applicable law. Do not expose it to untrusted input in an agentic setup without your own safeguards; with no refusal behaviour, it will not push back on injected instructions.

Quantization

The oQ5e weights were quantized directly from the abliterated BF16 source. The sensitivity model was used only to guide mixed-precision allocation; its quantized weights are not the source of this model. For this revision it is the previous oQ5e revision of this repository, which has the same abliterated lineage as the source; the previous revision itself used jedisct1/Qwen3.8-Flash-Next-Uncensored-oQ4e-100K-MTP. Nominal group size was 64; oQ mixed-precision may use different effective settings for selected tensors.

What changed in this revision

The previous revision was quantized with oMLX 0.6.4; this one with oMLX 0.7.3. Same source model and same oQ5e level. Besides the quantizer, the sensitivity model changed: the previous revision used the jedisct1 oQ4e build, this one uses the previous oQ5e revision of this repository. Between the two oMLX releases the following parts of oQ changed and apply to this model:

• Calibration of the importance matrix (imatrix). oMLX 0.7.x calibrates one layer at a time (layer-streaming), with a dedicated path for the Qwen4-Exp architecture, instead of needing the whole model, or a 4-bit proxy copy of it, to fit in memory. For MoE models it also keeps sampling until the experts have seen enough tokens (128 sequences at a time, up to 1,024). For this build the report (oq_imatrix_report.json) shows 1,024 sequences of 512 tokens, 937 imatrix entries, and 75,240 of 75,264 routed experts (99.97%) activated at least once. 24 experts received no calibration tokens, so the adaptive coverage target was not fully met.
• Group scales and offsets refit by weighted least squares. After the importance-weighted search, oQe now refits each group's scale and offset a few times and keeps the refit only if it lowers the weighted error. The oMLX authors report a 12% to 15% lower KL divergence from the original model on Qwen3.5-9B and Qwen3.6-35B-A3B; this was not measured on this model.
• Bit allocation. The output head and embeddings are sized from the base width and return to 8-bit when budget remains (both are 8-bit here). The MTP head is now planned after the rest of the model, so keeping it does not change the bits of the other layers; it keeps at least 4-bit.
• Engine, not weights. oMLX 0.7.x also contains many Qwen3.8-Flash-Next decode, prefill and MTP-verify speedups. They live in the engine and apply to any copy of the weights run on oMLX 0.7.3; they are not a property of this quantization.

On MMLU and MMLU-Pro the new revision scores 88.1% and 70.9%, against 88.0% and 70.4% for the previous one (see Benchmarks). Those differences are within sampling noise, so this revision should be read as a re-quantization with an improved pipeline and equal measured quality, not as a measurable accuracy gain.

Hardware and memory

Qwen3.8-Flash-Next contains a very large N-gram/PLE component. With SSD N-gram Offload it does not all stay resident in unified memory, which is what makes a model this size practical on a 128 GB machine. Long-context and KV-cache use raise memory further with context length.

Tested on a MacBook Pro, Apple M4 Max, 128 GB, oMLX, with SSD N-gram Offload enabled. Peak MLX allocation (weights + KV + activations), measured in oMLX with SSD N-gram Offload on: ~88 GB at 4K context, ~92 GB at 128K. These are process-level allocator peaks, not total-system memory. They were measured on the previous revision and not re-measured for this one.

Recommended oMLX settings

For a 128 GB Apple Silicon Mac:

• SSD N-gram Offload: ON
• Lightning MTP: ON for chat/reasoning; consider OFF for coding agents

Serve via oMLX's OpenAI-compatible API. The weights are standard mlx-lm compatible safetensors and should also load in mlx-lm and other MLX apps (untested).

Note: an unrelated oMLX engine bug on 0.7.0.dev1/dev2 can silently drop tool calls to unregistered function names on the streaming path (jundot/omlx#3660, open as of 2026-09; 0.6.4 unaffected). It does not affect the weights.

Benchmarks

Practical local tests, limited by compute time, not a standardized academic suite. The oQ5e [this repo, current] column is this revision (oMLX 0.7.3); the other columns were measured earlier, on oMLX 0.7.0.dev2.

Thinking OFF

¹ jedisct1/Qwen3.8-Flash-Next-Uncensored-oQ4e-100K-MTP (abliterated lineage) ² GBP-DE/Qwen3.8-Flash-Next-oQ5e-mtp (base, non-abliterated)

At 2000 samples MMLU is flat across all builds (~88%), within sampling noise: the abliterated quants are at parity with the base build on broad knowledge. On MMLU-Pro, accuracy rises with bit width across the oQ4e/oQ5e/oQ6e series. The current oQ5e revision is +0.1 points on MMLU and +0.5 points on MMLU-Pro over the previous one; at 1000 samples the MMLU-Pro standard error is about 1.4 points, so this is not a significant difference.

Smaller (n=100) GSM8K and HumanEval runs were near ceiling (90-98%) for every build and are omitted as non-discriminating.

Model architecture

Quantized from Qwen3.8-Flash-Next (sparse Mixture-of-Experts): ~125B LM parameters, ~6B activated, ~51B N-gram embedding parameters, ~4B MTP parameters, 48 layers, 512 experts, 10 routed + 1 shared activated. See the official Qwen model card for full architecture and context details.

Credits

• Base (abliterated) model: orcarouter/Qwen3.8-Flash-Next-Uncensored, itself an abliteration of Qwen/Qwen3.8-Flash-Next.
• Quantization: oMLX 0.7.3, oQ5e enhanced/imatrix (previous revision: oMLX 0.6.4).
• Sensitivity model: the previous oQ5e revision of this repository (the previous revision used jedisct1/Qwen3.8-Flash-Next-Uncensored-oQ4e-100K-MTP).

License

Qwen Community License 1.0, inherited from the source repository and included as LICENSE. Note the source card labels itself Apache-2.0, which does not match the license file it ships; review before use or redistribution.