ThinkingCap-Qwen3.6-27B-3bit-mlx

leonsarmiento/ThinkingCap-Qwen3.6-27B-3bit-mlx

This model was converted to MLX format from bottlecapai/ThinkingCap-Qwen3.6-27B using mixed 3-bit quantization optimized for Apple Silicon. Vision tower stripped for text-only inference.

Quantization Details

• Bits per weight: 3.596
• Total size: ~11 GB (3 shards)
• Method: mlx_lm.convert with custom quant_predicate
• Note: Same recipe as leonsarmiento/Qwen3.6-27B-3bit-mlx

About ThinkingCap

ThinkingCap-Qwen3.6-27B is a token-efficient thinking model fine-tuned from the Qwen3.6-27B base, optimized for efficient reasoning with reduced token usage during thinking.

Intelligence Benchmarks (Instruct Mode)

Compared against other Qwen3.6-27B MLX variants in the same bit-depth range. All benchmarks run in instruct mode (thinking disabled), n=30 per benchmark.

Thinking-Mode Benchmark (Custom MATHQA, n=10)

The instruct-mode benchmarks above do not measure ThinkingCap's actual value proposition — it is a token-efficient thinking fine-tune. To test what it was built for, a custom MATHQA sample (10 random questions from our curated 30-question set, seed=42) was run on four Qwen3.6-27B variants in thinking mode with an 8192-token budget.

¹ Average reasoning chain length (characters) on correct answers only — lower is more efficient.

Per-Question Breakdown

² Bonsai Q1_0 hit the 8192-token limit in a degenerate self-verification loop without producing a clean answer letter. Q379: 12,887 reasoning chars, Q26: 9,730 chars, Q383: 10,937 chars.

Analysis

ThinkingCap's value proposition is fully validated in thinking mode. The instruct-mode ranking inverts completely:

• +10 points accuracy over Ternary-Bonsai and oQ4e-mtp (90% vs 80%), +20 points over Bonsai Q1_0 (90% vs 70%). ThinkingCap is the only model to solve Q379 (a probability question that both Ternary-Bonsai and oQ4e-mtp get wrong in all modes).
• 36% less reasoning tokens than Ternary-Bonsai on correct answers (2,555 vs 3,962 chars), and 70% less than Bonsai Q1_0 (8,544 chars). The "token-efficient thinking" fine-tune produces tighter, more decisive reasoning chains — exactly as designed.
• On simple problems (Q418, Q914, Q653), ThinkingCap uses 3–6× less reasoning than Ternary-Bonsai while reaching the same correct answer. On Q418: 918 chars vs 5,901 chars.
• Q383 is universally hard — all four models fail it. ThinkingCap spirals into a 14,971-char degenerate loop (its one instance of overthinking). Bonsai Q1_0 degenerates on 3 questions (Q379, Q26, Q383), all hitting the 8192-token limit. Ternary-Bonsai degenerates on 2 questions (Q379: 11,175 chars output, Q383: 6,413 chars).
• Bonsai Q1_0 (1-bit) is the smallest model (3.8 GB) but pays a steep quality cost: lowest accuracy (70%), worst token efficiency (8,544 chars/correct — 3.3× more reasoning than ThinkingCap), and 3 degenerate loop failures. It also runs the slowest (195.7s avg) due to the massive reasoning chains. Even on questions it gets right, it uses 4–6× more reasoning chars than ThinkingCap (e.g., Q653: 12,441 vs 1,116, Q2619: 13,872 vs 1,441).
• oQ4e-mtp is the throughput king (62.9s avg vs 111–114s for Ternary/ThinkingCap, 196s for Bonsai Q1_0) thanks to its 4-bit precision, but its thinking quality matches the 2-bit Ternary-Bonsai, not ThinkingCap.

Bottom line: ThinkingCap is worth keeping on the SSD if you ever use thinking mode. Its instruct-mode benchmarks understate its real-world value — in the reasoning workload it was optimized for, it is the clear winner on both accuracy and token efficiency, despite being a 3-bit model competing against a 4-bit. The 1-bit Bonsai Q1_0, while impressively small at 3.8 GB, demonstrates that extreme quantization causes severe overthinking degeneration — its reasoning chains are long and circular rather than tight and decisive.

Test methodology: 10 random questions (seed=42) from the curated 30-question MATHQA set originally built for the Gemma 4 analysis. Thinking mode enabled, 8192 max tokens, temperature 1.0, top_p 0.95. Same prompt and parameters for all models. Bonsai Q1_0 served via llama.cpp at localhost:2727; all MLX variants via LM Studio at localhost:4321. Full results: gemma-4-results/qwen36_27b_thinking_mathqa.json, gemma-4-results/bonsai_q1_thinking_mathqa.json.

Recommended Inference Parameters - Add to Jinja template on LM studio or Chat Template Kwargs on oMLX

Thinking Preserve ({%- set preserve_thinking = true %}):

• General tasks: temperature=1.0, top_p=0.95, top_k=20, min_p=0.0, presence_penalty=1.5, repetition_penalty=1.0
• Coding tasks: temperature=0.6, top_p=0.95, top_k=20, min_p=0.0, presence_penalty=0.0, repetition_penalty=1.0

Instruct Mode ({%- set enable_thinking = false -%}):

• General tasks: temperature=0.7, top_p=0.8, top_k=20, min_p=0.0, presence_penalty=1.5, repetition_penalty=1.0
• Reasoning tasks: temperature=1.0, top_p=0.95, top_k=20, min_p=0.0, presence_penalty=1.5, repetition_penalty=1.0

Use with mlx

pip install mlx-lm

from mlx_lm import load, generate

model, tokenizer = load("leonsarmiento/ThinkingCap-Qwen3.6-27B-3bit-mlx")

prompt = "hello"
messages = [{"role": "user", "content": prompt}]
prompt = tokenizer.apply_chat_template(messages, add_generation_prompt=True, return_dict=False)

response = generate(model, tokenizer, prompt=prompt, verbose=True)