ThinkingCap-Qwen3.6-27B-Alis-MLX-Dynamic

Apple Silicon (MLX) mixed-precision quantization of bottlecapai/ThinkingCap-Qwen3.6-27B — the token-efficiency finetune of Qwen3.6-27B (~50% fewer thinking tokens at preserved accuracy). 4.601 bpw, 15.47 GB, GPQA-Diamond 83.3% — 0.5pt from bf16.

build	weights	GPQA-D*	avg think tokens	decode†	peak mem
main (this) — multimodal	16.13 GB	**83.3%**¹	2,838	35.6 tok/s	16.7 GB
text — text-only	15.47 GB	83.3%	2,838	36.5 tok/s	15.8 GB
BottleCapAI Q4_K_M GGUF (official)	16.81 GB	82.3%	3,654	—	—
BottleCapAI bf16 (original)	55.56 GB	83.8%	3,019	—	—

* n=198, temperature 1.0, seed-fixed shuffled choices, 32K token cap — identical harness for all three rows (harness validated by reproducing the ThinkingCap card's published 83.8 exactly). † single-stream on an M3 Ultra, short prompt.

Why this build

Smaller and more accurate than BottleCapAI's official 4-bit GGUF, with 22% fewer thinking tokens per answer — on Apple Silicon, fewer generated tokens is the speedup that matters. Fits 24 GB Macs with room for context (15.8 GB measured peak at short context; int8 KV keeps long-context overhead small — only 16 of 64 layers carry KV in this 3:1 DeltaNet hybrid).

The finding behind it: short thinkers survive quantization

Across our GPQA campaign, the same 4.6bpw recipe loses 12.8pt on vanilla Qwen3.6-27B (85.5 → 72.7) but only 0.5pt on ThinkingCap (83.8 → 83.3). Long reasoning chains accumulate quantization drift token by token; ThinkingCap's ~3K-token chains stay on track where vanilla's ~11K-token chains derail. Token-efficiency finetune first, quantize second — the two compound.

Multimodal (default)

main carries the full multimodal stack: the 4.6bpw text weights (¹ GPQA measured on the identical text weights) + the vision tower at 8-bit. 16.13 GB total — still smaller than the 16.81 GB text-only GGUF. Verified with OCR + visual-reasoning tests (EN/KO). Text-only users: main loads fine under plain mlx-lm (vision weights are skipped), or use the text branch to save 0.66 GB.

pip install mlx-vlm
python -m mlx_vlm generate \
  --model avlp12/ThinkingCap-Qwen3.6-27B-Alis-MLX-Dynamic \
  --image photo.jpg --prompt "이미지의 텍스트를 읽어줘." --max-tokens 300

Usage

from mlx_lm import load, generate

model, tokenizer = load("avlp12/ThinkingCap-Qwen3.6-27B-Alis-MLX-Dynamic")
prompt = tokenizer.apply_chat_template(
    [{"role": "user", "content": "뮤텍스와 세마포어의 차이를 설명해줘."}],
    add_generation_prompt=True,
)
print(generate(model, tokenizer, prompt=prompt, max_tokens=512))

mlx_lm.generate --model avlp12/ThinkingCap-Qwen3.6-27B-Alis-MLX-Dynamic --prompt "..." -m 512
mlx_lm.server --model avlp12/ThinkingCap-Qwen3.6-27B-Alis-MLX-Dynamic --port 8080

Long context: --kv-bits 8 --kv-group-size 64 (measured quality-free on this family: KL vs bf16-KV ≤ 0.003).

Recipe (verified from config.json)

Per-tensor bits by marginal utility from a full weight-space sensitivity scan, validated end-to-end (the scan alone misranks formats — see the nvfp4 note on the vanilla 27B repo):

component	precision
bulk (90.1% of params)	affine 4-bit g64
sensitivity-ranked top 9.8%	affine 5-bit g64
routers / small gates	6-bit pins
embed / lm_head	≥4-bit

No DWQ retune: at ≥4bpw the validation-gated DWQ rejects every round (already teacher-close); conversion-time weights ship as-is.

Notes & provenance

• main carries the full multimodal stack (mlx-vlm); the text branch is the mlx-lm text-only artifact.
• Eval artifacts: same 198-question GPQA protocol for every row of the table above.
• Built on an M3 Ultra 512 GB · mlx-lm 0.31.3 · mlx 0.31.2 · 2026-07.
• Siblings: Qwen3.6-27B-Alis-MLX-Dynamic (vanilla golden + 5.5bpw reasoning tier), Qwen3.6-35B-A3B-Alis-MLX-Dynamic (MoE, 102 tok/s).

Credits

• bottlecapai — ThinkingCap-Qwen3.6-27B, the token-efficiency finetune this quantizes (and the Q4_K_M GGUF benchmarked against).
• Qwen team — the underlying Qwen3.6-27B.
• Apple MLX team — mlx / mlx-lm.
• Quantization recipe & evals: Alis (avlp12) with the alis-dwq pipeline; built with assistance from Claude (Anthropic).