Qwen3.8-27B-oQ8e-fp16-mtp

Qwen3.8-27B-oQ8e-fp16-mtp

8-bit MLX quantization of Qwen/Qwen3.8-27B (hybrid Gated DeltaNet + Gated Attention, native vision, 262K context, thinking mode with preserved reasoning). Made with oQ (oMLX v0.6.0.dev1) mixed-precision quantization.

Sibling repos:

• Qwen3.8-27B-oQ4e-fp16-mtp — 4-bit / fp16, ~16.7 GiB, tuned for M1/M2; first ANE-eligible build of the family (oMLX 0.6.1, +21% prefill on M2 Ultra). As of oMLX 0.6.3 this 8-bit build is ANE-eligible too — see the measured section below.
• Qwen3.8-27B-oQ6e-mtp — 6-bit / bf16, ~22 GB, for 64 GB-class Apple Silicon
• Qwen3.8-27B-oQ4e-mtp — 4-bit / bf16, ~17 GB, for 32 GB-class machines or long-context use on 64 GB
• Qwen3.8-27B-oQ8e-mtp — 8-bit / bf16, ~27.9 GiB; same quantized tensors and imatrix as the oQ8e-fp16 build, bf16 floating tensors for M3/M4+ numerical safety

Quantization details

• Model type: qwen3_5
• Bits: 8 (effective 8.6 bpw mixed precision), group size 64 — ~27.5 GiB weights
• Enhanced quantization (oQe): imatrix-calibrated (1024 samples) — affine quantization weighted by activation importance
• MTP weights preserved (mtp.* tensors + config) — multi-token-prediction / Lightning MTP works after quantization
• Non-quant weight dtype: float16 — ~20% faster prefill on M1/M2 Apple Silicon (native fp16). On M3/M4, prefer the bf16 sibling repo for numerical safety.
• Vision components included (not a text-only strip)
• Format: MLX safetensors

Recommended sampling (per the Qwen3.8 model card)

Thinking controls via chat_template_kwargs:

• enable_thinking (default true)
• preserve_thinking (default true) — keeps reasoning traces across multi-turn history
• reasoning_effort: xhigh (default) / medium / low — in our testing, medium reduced thinking volume ~25% with no loss on agentic tasks

Tested

Validated 2026-08-14 on an M2 Mac Studio (this exact quant), served via an OpenAI/Anthropic-compatible gateway:

• Multi-turn agentic tool use with thinking ON and thinking blocks fed back into history: 0 stalled turns across a 3-scenario synthetic battery (~8 tool-call turns) and a 12-turn live session reaching 27 messages of history — tool-call emission stayed reliable throughout. This is a notable improvement over Qwen3-generation models, where returning thinking blocks in history degraded tool calling; Qwen3.8's preserved-thinking design handles it.
• Reasoning depth flexes with task difficulty (observed ~100 chars of thinking on routine turns, ~4,000 on decision-heavy turns).
• Thinking mode costs roughly 25–35% wall time vs. instruct mode on agentic workloads.
• Clean reasoning_content separation on /v1/chat/completions; proper thinking + tool_use blocks over an Anthropic-style /v1/messages bridge.
• Context: validated at the full native 262,144 on the M2 Mac Studio (oMLX built-in context test); client deployments use a 229,376 window (262,144 − 32,768 output headroom). On smaller machines, prefill-time allocation — not KV — is the binding constraint; see the 4-bit sibling's card for measured 64 GB numbers.

Measured: ANE prompt processing (oMLX 0.6.3)

oMLX 0.6.3 extended its experimental Qwen ANE prefill path to q6/q8 layouts (bit-aware affine, replacing the 0.6.1/0.6.2 INT8 approximation that was q4-only), so this build is now ANE-eligible — earlier revisions of this card correctly called it ineligible; that ended with 0.6.3. The 0.6.3 path also adds an FP16 CPU branch (AMX work sharing), which this fp16 variant feeds natively.

Measured on an M2 Ultra Mac Studio (128 GB), oMLX 0.6.3, built-in benchmark (Code/Python context, ANE-aligned prompts off — i.e. realistic ragged prompt lengths), one run per leg, same day, identical settings apart from ANE:

Split (found by the 0.6.3 Tuner V2, dual-ANE machine): MLP 50% on ANE, GDN 37.5% on ANE (48-layer limit), CPU shares 10% gate / 20% down-projection / 10% GDN with 8 workers, tail padding from 1,436 tokens. The tuner's own shape measured +42.7% (438.3 pp tok/s); the gain compresses with depth as the non-ANE full-attention layers claim a growing share of prefill — same pattern as the 4-bit sibling.

Notes before enabling:

• Prompts at or under ~1K gain nothing — they never fill the 2,048-token ANE program block and ride the GPU path unchanged.
• Decode is untouched at depth (26.6 → 26.4 tg tok/s at 64K). One watch item: our single 1K leg showed ~8% slower decode with ANE resident — beyond typical ±3.5% run variance but unreplicated; re-measure on your machine if short-turn latency matters.
• Cost: ~16 GB extra resident memory for the model's lifetime (measured peak 44.0 → 60.0 GB at 64K) — more than the 4-bit sibling's ~11 GB on the same path. On machines without clear headroom above the model + KV, leave it off.
• The gain applies to cold prefills only; prefix-cached turns are unaffected, and 0.6.3 itself improved prefix-cache reuse in long agentic sessions.
• The first request after a load carries extra warm-up on top of the eager ANE compile; subsequent requests match steady state.
• Quality-checked with ANE on (2026-08-27, this machine): the family's multi-turn agentic battery passed 3/3 scenarios, 0 stalled turns across 8 tool-call turns, thinking fed back into history verbatim. The battery is short-context; 0.6.3's recurrent-safe ANE design (token-local rows only on the approximate path) addresses the earlier long-prompt divergence reports upstream.