license: apache-2.0
base_model:

• AEON-7/Qwen3.8-27B-AEON-ULTIMATE-UNCENSORED-BF16
  base_model_relation: quantized
  library_name: mlx
  pipeline_tag: image-text-to-text
  language:
• en
• zh
• multilingual
  tags:
• mlx
• mtplx
• qwen3.8
• mixed-precision
• affine-quantization
• speculative-decoding
• multi-token-prediction
• vision
• tool-calling
• reasoning
• uncensored

AEON Qwen3.8 27B — Attention8 + BF16 Recurrence for MLX/MTPLX

A quality-first Apple Silicon conversion of
AEON-7/Qwen3.8-27B-AEON-ULTIMATE-UNCENSORED-BF16.
It packages the mixed-precision language model, BF16 vision tower, tokenizer,
tool template, and 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.

At a glance

Use this release when you have a 64 GB-or-larger Apple Silicon Mac and want the
quality-focused local build. If memory or storage is tighter, use the
16.08 GB MXFP4 sibling.
The 29.53 GB MXFP8 sibling
uses a larger, simpler precision strategy. Those siblings have not yet received
the same final benchmark invocation, so size does not imply a quality ranking.

Why this hybrid is larger than a 4-bit model

The quantization recipe spends precision where local regression tests found the
architecture most sensitive:

• 168 language modules use affine 4-bit weights with group size 32.
• 234 attention, embedding, output, and late-MLP modules use affine 8-bit
  weights with group size 64.
• 96 recurrent in_proj_a and in_proj_b projections remain BF16.
• The vision tower and matching 15-tensor MTP head remain BF16.

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

Qualification snapshot

The recipe-selection suite chose this hybrid over the compact local 4-bit
baseline: hard-task accuracy improved from 6/10 to 8/10, repeated code
checks remained 6/6, and code decode stayed close at 48.10 versus
50.56 tok/s. Full candidate comparisons are in BENCHMARKS.md.

MTPLX 2.9 performance retest

A performance-only retest of the exact published weight-bearing revision
b5a54ea5d7745b6ddada238f83b66d63c979b9a5
used MTPLX 2.9.0 on the same Apple M4 Max 64 GiB host:

The API code decode result is 14.0% above the original MTPLX 2.7.1
recipe-selection result of 48.10 tok/s. D3 remained the winner with 90.76%
aggregate draft acceptance in the isolated lane. Testing used automatic fans;
no privileged fan pinning or 80–90 tok/s claim is included. The isolated lane
used the recommended temperature 1.0/top-p 0.95/top-k 20 target and draft
samplers. The API code lane was a deterministic temperature-0, thinking-off
diagnostic with unquantized KV cache.

This retest updates performance evidence only. The complete quality, context,
tools, agentic, and vision qualification below remains the MTPLX 2.7.1 run.
See BENCHMARKS.md and the
MTPLX_2.9_PERFORMANCE.json sanitized aggregate.

Post-publication Shiftedx Bench gate

• Evaluated model revision: c839d980679e9aec1ab66388ded0689040c9d93a
• Benchmark revision: Shiftedx Bench v0.3.0 3bbb0bfa
• Runtime: MTPLX 2.7.1 turbo, MTP depth 3, thinking on/medium, temperature
  1.0, top-p 0.95, top-k 20, unquantized KV cache
• Host: Apple M4 Max with 64 GiB unified memory

Scores are reported separately; Shiftedx Bench does not produce a composite
“intelligence” score. The sanitized gate artifact is
SHIFTEDX_BENCH_V0.3_TOOLS_AGENTIC_VISION.json.

Expanded agentic comparison

The two-case gate above is a regression check, not a broad agent-readiness
claim. Shiftedx Bench v0.5.1 ran three complete paired trials with and without
the optional shiftedx-harness-v1 execution policy:

The harness converts observable tool receipts into bounded execution state. It
does not change the weights or see hidden answers/tests. On the expanded matrix
it reduced dispatched tool calls from 282 to 267 but increased total wall time
by 11.4%. Repository trials were more variable; baseline won one of the three.
Treat this as a higher-reliability deployment policy at a different operating
point—not a weight-level intelligence improvement or a free speedup.

• Evaluated model revision: b5a54ea5d7745b6ddada238f83b66d63c979b9a5
• Benchmark revision: Shiftedx Bench v0.5.1 335e6694
• Readable report, family scores, telemetry, sanitized per-case rows, and manifests

Quick start

Install MTPLX 2.9.0,
the release used for the performance retest above, then:

MODEL=Shiftedx/qwen3.8-27b-aeon-ultimate-uncensored-attention8-bf16recurrence-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. For the
published 6/6 tool result, ordinary/multi-round requests used the hybrid tool
contract while explicit parallel intent used Qwen's native template. That
router policy is included as
MTPLX_2.7.1_QWEN38_EXPLICIT_PARALLEL_NATIVE.patch.

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.

Context and memory

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

Reaching the full window may require a larger-memory Mac or a separately
qualified q8/q4 KV-cache configuration. Changing KV precision creates a
different runtime candidate and requires new quality, context, performance, and
stability measurements. A 32 GiB Mac was not tested and may be too constrained;
prefer the MXFP4 sibling when memory is tight.

Provenance and integrity

• Source: AEON-7/Qwen3.8-27B-AEON-ULTIMATE-UNCENSORED-BF16
• Pinned source revision: a6775a9a8ebb65cab3f707b4ab087fc7aa698634
• BUILD_RECIPE.json: precision policy
• mtplx_runtime.json: native-MTP contract and defaults
• RELEASE_MANIFEST.json and SHA256SUMS: file integrity
• AEON Qwen3.8 MLX collection

The vision tower and MTP sidecar are preserved from the qualified 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.
• Tool calls, JSON, code, image interpretations, and factual answers can be
  wrong. Validate output before allowing system changes or code execution.
• Only the Apple M4 Max 64 GiB configuration received the complete published
  qualification.
• 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.