Qwen3.8-27B-AEON-ULTIMATE-UNCENSORED-MLX-MIXED

The Apple Silicon seat in the AEON ULTIMATE UNCENSORED family. A mixed-precision MLX conversion of AEON-7/Qwen3.8-27B-AEON-ULTIMATE-UNCENSORED-BF16, packaging the quantized language model, an untouched BF16 vision tower, the tokenizer and tool template, and the native BF16 multi-token-prediction head in one MTPLX-ready repository.

This is an independent community conversion, not an official AEON-7, Qwen, Apple, MLX, or MTPLX release.

Model family (pick your seat)

The parent publishes a BF16 master and an NVFP4 deploy cut. This is the third seat: the same will, quantized for a Mac instead of a datacenter card.

At a glance

Why "MIXED"

The name is deliberate, and it means the same kind of thing the parent's NVFP4-MIXED means: precision is not uniform. It is spent where local regression tests found this architecture most sensitive, and saved everywhere else. The lattice differs from the parent's because the numeric formats differ — this is MLX affine quantization, not ModelOpt NVFP4.

Two structural choices carry most of the quality:

• The recurrence stays in BF16. Qwen3.8's Gated DeltaNet layers carry state through in_proj_a and in_proj_b. Quantization error there compounds along the sequence instead of averaging out, so all 96 of those projections are left at source precision. This is the single most important line in the recipe.
• The last eight layers keep full 8-bit MLPs. Layers 56–63 do disproportionate work shaping the output distribution, which is the same reason the parent's NVFP4 lattice holds its last-8 MLPs at FP8.

This is a custom MLX affine mixed-precision model — not MXFP4, MXFP8, NVFP4, GGUF, AWQ, or GPTQ. The exact module map is in BUILD_RECIPE.json. Do not infer quality from file size.

Early Access Draft

The parent labels its BF16 master an Early Access Draft, not GA, and that label passes through to this conversion. Quantizing a draft does not promote it.

What the parent says about the draft, and what you should expect here:

• On ordinary work it behaves like a 27B. The refusal layer is gone; you can ask a hard question and get an answer.
• On very long generations, small gaps in the abliterated weights stack up. A loop can appear, a phrase can repeat, a thought can fail to land. That is the draft, not the quantization — and not your imagination.

A GA cut of the source is in progress upstream: the next evolution of abliteration for Qwen3.8, aiming to keep the instructional structure while changing the will. When that GA checkpoint lands, it will be quantized to this same MLX-MIXED recipe and published as a sibling release. This repo will stay pinned to the Early Access lineage so anything built against it keeps working.

Quick start

Install MTPLX 2.11.2 or newer, then:

MODEL=Shiftedx/Qwen3.8-27B-AEON-ULTIMATE-UNCENSORED-MLX-MIXED-VISION-MTPLX

mtplx pull "$MODEL"
mtplx inspect "$MODEL" --require-mtp
mtplx start cli \
  --model "$MODEL" \
  --profile turbo \
  --depth 3 \
  --reasoning on \
  --reasoning-effort medium \
  --tool-prompt-mode native \
  --chat-template-profile tokenizer

For an OpenAI-compatible endpoint:

mtplx serve \
  --model "$MODEL" \
  --download \
  --host 127.0.0.1 \
  --port 8000 \
  --profile turbo \
  --generation-mode mtp \
  --load-mtp \
  --depth 3 \
  --reasoning on \
  --reasoning-effort medium \
  --tool-prompt-mode native \
  --chat-template-profile tokenizer \
  --paged-kv-quantization off

MTPLX accepts OpenAI tools and standard image_url content parts. Recommended starting point: turbo profile, MTP depth 3, thinking on at medium effort, temperature 1.0, top-p 0.95, top-k 20, and unquantized KV cache. Run mtplx tune --model "$MODEL" --retune on your own Mac, because acceptance and speed vary with chip, context, thermals, and workload.

The source's chat template supports three reasoning efforts — low, medium, and xhigh (the template default). This release is tuned and recommended at medium. To suppress reasoning for a low-latency request, send the top-level OpenAI field "enable_thinking": false.

Qualification

Measured on MTPLX 2.11.2

A depth sweep against autoregressive decoding on an Apple M4 Max 64 GiB, turbo profile, fans on auto (no privileged fan pinning, no 80–90 tok/s claim):

Every candidate passed its runtime and quality gate; depth 3 won outright. The same sweep under the sustained profile reached 54.93 tok/s at D3, so turbo is worth 11.0% here — expected, since turbo adds verify-specialized quantized-matmul kernels and a compiled verify step for 4-bit bodies, and this body is 168 affine-4 modules.

Gates cleared on this build: mtplx inspect verified native runtime contract, cold-start MTP inference at depth 3, a text smoke, and a vision smoke that named all three colour/shape pairs in a synthetic test image.

Throughput is host- and workload-specific. Run mtplx tune on your own Mac.

Relationship to the earlier attention8 release

This artifact is bit-identical to the earlier attention8 + BF16 recurrence release. All eleven body shards, the vision shard, and the MTP sidecar hash the same SHA-256, verified against that release's published RELEASE_MANIFEST.json.

That is not a coincidence and it is worth stating plainly. The source repository was revised between the two builds, but its weight files did not change — only the model card, .gitattributes, and tokenizer_config.json, which now inlines the chat template that already shipped as chat_template.jinja. Rebuilding the same recipe on a newer toolchain (mlx 0.32.2, MTPLX 2.11.2 rather than 0.32.1 and 2.7.1) reproduced the quantization exactly.

Two consequences:

• The earlier release's quality qualification — 8/10 on its seeded hard suite, 6/6 repeated code, 6/6 tool schema, vision pass, and exact-token context through 131,072 prompt tokens — describes these bytes, not merely a similar build. It is carried forward rather than re-measured.
• Only the performance numbers above are new, because only the runtime changed.

This release differs from that one in name, in shipping the source's updated tokenizer_config.json, and in carrying a runtime contract verified under MTPLX 2.11.2 at the turbo profile.

Context and memory

The model advertises a 262,144-token context window. The bit-identical earlier release passed exact-token qualification through 131,072 prompt tokens with an unquantized KV cache; a 260,096-token attempt exceeded the practical memory envelope of the 64 GiB reference host. That is reported as not qualified on this host/runtime, not as a wrong answer.

Peak memory during the text smoke was 22.32 GB. Reaching the full window may require a larger-memory Mac or a separately qualified quantized KV-cache configuration; changing KV precision creates a different runtime candidate and needs its own measurements. A 32 GiB Mac was not tested and is likely too constrained.

Provenance and integrity

• Source: AEON-7/Qwen3.8-27B-AEON-ULTIMATE-UNCENSORED-BF16
• Pinned source revision: 0a197319321945a4f254e3dff616805d49d61206
• Source release stage: Early Access Draft
• BUILD_RECIPE.json: precision policy
• mtplx_runtime.json: native-MTP contract and defaults
• RELEASE_MANIFEST.json and SHA256SUMS: file integrity

The vision tower and MTP sidecar are preserved from the source lineage. Model behavior still depends on the exact runtime, template, sampler, reasoning mode, MTP depth, KV precision, and hardware.

Limitations and safety

• Quantization can change outputs relative to BF16; qualification does not guarantee source parity on every prompt.
• The source is an Early Access Draft with a known long-generation looping edge.
• Tool calls, JSON, code, image interpretations, and factual answers can be wrong. Validate output before allowing system changes or code execution.
• This intentionally uncensored source lineage may produce harmful, illegal, offensive, or unsafe material. Quantization does not restore alignment.

Operators are responsible for prompts, outputs, authentication, rate limits, moderation, audit logging, and human review appropriate to their deployment. Do not treat output as professional legal, medical, financial, or security advice.

License and credits

Apache-2.0, inherited from the source lineage. Follow the source model's license and attribution requirements when redistributing or deploying this derivative.

Credits to Qwen, AEON-7, MLX, MLX-LM, and MTPLX.