libertydesignstudio/OverSeer-Qwen3.8-27B-MTPLX

OverSeer Qwen 3.8 27B MTPLX

A mixed-precision MLX build of Qwen/Qwen3.8-27B with its native MTP draft head preserved for speculative decoding through MTPLX.

The release target is a 24 GB Apple Silicon workstation: approximately the memory footprint of a Q3_K_M GGUF while materially improving decode speed.

Measured result

Tested on a base Apple M4 with 24 GB unified memory, macOS 26.5.2, MLX 0.32.0, mlx-lm 0.31.3, and MTPLX 2.7.2.

Measurement	Result
Loaded model weights	13.4 GB
Forge D2 verification	13.79 tok/s
OpenAI server, two 300-token prompts	11.23–14.01 tok/s
Long reasoning samples	10.95–13.08 tok/s
Same-artifact AR baseline	7.41 tok/s
Prior LM Studio Ridge baseline	5.26–5.79 tok/s

The server measurements used performance-cold, MTP depth 2, one active request, a one-entry/1 GB session bank, and extended background warmup disabled. Generation speed excludes prompt prefill; end-to-end latency still depends on prompt length.

Quality checks

The release candidate was checked on three paired tasks:

• Levenshtein edit-distance implementation: executable code passed five deterministic cases; stated O(m·n) time and space.
• TCP versus UDP explanation: correct connection, reliability, ordering, and analogy requirements.
• First-occurrence binary-search repair: executable code passed seven boundary and duplicate cases.

Known limitation: an optional illustrative edit-distance table contained incorrect intermediate cells, although both generated implementations, the complexity analysis, and the final distance were correct. The debugging task needed a 4,096-token reasoning allowance to reach its correct final answer; 1,800 tokens ended during reasoning. This is a small release check, not a broad quality benchmark.

Quantization

The body defaults to 2-bit affine quantization with higher precision on quality-sensitive paths:

• embeddings and LM head: 6-bit
• linear-attention QKV/Z/out: 4-bit
• linear-attention A/B: 8-bit
• full-attention Q/K/V: 5-bit; output: 6-bit
• all MLP down projections: 3-bit
• first and last four MLP gate/up projections: 3-bit
• native MTP sidecar: BF16

The complete machine-readable recipe is in quantization_recipe.json.

Forge verification patch

The included MTPLX 2.7.2 patch is needed to reproduce Forge's MTP-depth verification for this 2-bit body. The recommended server intentionally leaves MTPLX_NAX_VERIFY unset: enabling that experimental dispatcher reduced decode on an 8.1K-token OMP prompt from 10.23 to 2.94 tok/s. Stock MLX quantized matmul is the validated server lane.

python -m pip install 'mtplx==2.7.2'
SITE_PACKAGES="$(python -c 'import site; print(site.getsitepackages()[0])')"
patch -p1 -d "$SITE_PACKAGES" < mtplx-2.7.2-q2-verify.patch

The patch adds numerically checked 2-bit affine verify kernels for MTP row counts 2 through 6. Synthetic BF16 comparisons against stock MLX measured maximum absolute differences from 0.00225 to 0.00384 across those row counts.

Recommended text server

MTPLX_SESSION_BANK_MAX_ENTRIES=1 \
MTPLX_SESSION_BANK_MAX_BYTES=1G \
MTPLX_SESSION_BANK_PER_SESSION_BYTES=1G \
MTPLX_SESSION_BLOCK_PREFIX_RESTORE=0 \
MTPLX_MEMORY_PRESSURE_GUARD=0 \
MTPLX_WARMUP_EXTENDED=0 \
mtplx serve \
  --model ./OverSeer-Qwen3.8-27B-MTPLX \
  --profile performance-cold \
  --depth 2 \
  --scheduler-mode serial \
  --batching-preset solo \
  --max-active-requests 1

performance-cold is the measured fast lane for text requests under 8K prompt tokens (10.95–14.01 tok/s). Native OpenAI function calling works on a compact 604-token direct request (one valid call at 10.96 tok/s), but the full OMP tool/rule prompt expands to roughly 23.7K tokens: performance-cold correctly blocks that unsafe prefill, while sustained did not complete the end-to-end OMP agent turn within 15 minutes. This release is therefore recommended as a fast text/coding model, not as the default full-tool agent. Upstream model metadata retains 262K context; the qualified local OMP text route advertises 16K.

Files

• mixed-precision MLX language-model shards
• mtp.safetensors: native BF16 MTP head
• model-vision.safetensors: restored vision tower
• mtplx_runtime.json: runtime contract and measured verification evidence
• quantization_recipe.json: reproducible Forge recipe
• mtplx-2.7.2-q2-verify.patch: optional Forge/tuning verification patch

License

Apache 2.0, matching the upstream Qwen3.8-27B release. This derivative is not affiliated with or endorsed by the Qwen team.