Qwen3.6-27B MTPLX Optimized Speed V2
The recommended 27B model for local coding and agent work with
MTPLX on Apple Silicon.
V2 is a hand-calibrated quantization of Qwen3.6-27B built for one job: fast,
high-quality coding on a Mac. It cuts quantization loss by 47%
compared to the original Optimized Speed release, and it pulls further ahead
the longer an agent session runs.
Quickstart
1brew install youssofal/mtplx/mtplx
2mtplx start
mtplx start recommends this model automatically on Macs with 32 GB of
unified memory or more, and offers smaller verified models below that.
Direct usage:
1mtplx pull Youssofal/Qwen3.6-27B-MTPLX-Optimized-Speed-V2
2mtplx run "hello" --model Youssofal/Qwen3.6-27B-MTPLX-Optimized-Speed-V2
3mtplx serve --model Youssofal/Qwen3.6-27B-MTPLX-Optimized-Speed-V2 --port 8000
mtplx serve exposes OpenAI and Anthropic compatible endpoints, so the model
works inside OpenCode, Claude Code, Cline, and other coding agents.
The numbers
Measured on an M5 Max against the original Optimized Speed release, same
engine, same prompts:
- KL divergence to the full-precision bf16 model: 0.072 vs 0.135, on an
86k-token held-out battery. 47% less quantization loss than the original.
- Draft acceptance at 17k tokens of agent context: 98 / 86 / 78% per draft
position, vs 73 / 50 / 33% for the original.
- Decode at that context: 45.7 vs 36.1 tok/s, about 27% faster.
- Peak speculative multiplier: 2.86x over the matched autoregressive
baseline, the highest we have measured on any 27B build.
These numbers are one story, not four. Speculative decoding is only as fast as
its drafts are accurate, and a trunk that stays this close to full precision
keeps its drafts landing deep into long contexts. Quality is the speed lever.
The tradeoff: V2 is 19.9 GB vs 16.4 GB for the original, so cold short chats
can be a few percent slower. Everything else is faster.
How it is built
Not a round-to-nearest conversion. The layout was tuned tensor by tensor
against the bf16 reference:
- Most weights are 4-bit with small groups, refit against real activation
statistics with error-weighted least squares instead of naive rounding.
- The tensors that hurt most at 4-bit were promoted by hand: embeddings, the
output head, convolution and attention output projections, and the late MLP
layers run at 8-bit, with the most sensitive layers kept in full bf16.
- The native multi-token-prediction head ships as a calibrated sidecar
(
mtp/weights.safetensors). No external draft model is needed.
The result lands at roughly 5.8 bits per weight. The goal was never the
smallest file. It is the best coding model that still fits comfortably in
unified memory.
About MTPLX
MTPLX is an MLX-native inference engine built around Qwen's native
multi-token prediction: the model drafts its own tokens and the engine
verifies them with exact probability-ratio acceptance, so output quality
matches non-speculative decoding at any temperature. It adds warm session
caching for fast agent round-trips, a native macOS app, and a one-line
install via Homebrew or pip install mtplx.
Which model to pick
- Coding, tool use, long agent sessions: this model.
- Tight on storage, mostly short chat:
Qwen3.6-27B-MTPLX-Optimized-Speed
is 3.5 GB smaller.
- Lower-memory Macs:
Qwen3.5-9B
and
Qwen3.5-4B
are verified for smaller memory tiers.
Verification
mtplx_runtime.json records the verified runtime contract:
- architecture:
qwen3-next-mtp, maximum MTP depth 3, turbo profile
- sampler: temperature 0.6, top-p 0.95, top-k 20
- KLD and perplexity measured against the bf16 teacher on an 86k-token
held-out battery
- verified hardware: Apple M5 Max, 128 GB unified memory
- served model id:
mtplx-qwen36-27b-optimized-speed-v2
Files
model-*.safetensors: calibrated hybrid-quantized model shardsmtp/weights.safetensors: native MTP sidecarmtplx_runtime.json: verified MTPLX runtime contract- tokenizer and processor files from the Qwen model family
Sources and attribution
