Solstice-AI/ThinkingCap-Qwen3.6-27B-mlx-6Bit

ThinkingCap-Qwen3.6-27B (MLX 6-Bit)

Apple Silicon 6-Bit Native Quantization of BottleCap AI's Loop-Mitigated ThinkingCap 27B

---

Executive Summary

Solstice-AI/ThinkingCap-Qwen3.6-27B-mlx-6Bit is an Apple Silicon-native 6-bit quantization of ThinkingCap-Qwen3.6-27B, the landmark reasoning-efficiency model created by BottleCap AI (co-founded by Tomáš Mikolov).

Designed specifically to eliminate the "thinking loop" pathology common to frontier Chain-of-Thought (CoT) models, ThinkingCap trains the network to recognize convergence boundaries and terminate intermediate verification once a mathematical or logical solution is proven. This slashes redundant internal monologue tokens by ~50%, halving response latency while preserving reference-grade reasoning accuracy.

Quantized in high-precision 6-bit affine format (group_size: 64) and powered by Solstice Labs' terminal-first Anvil runtime, this release provides lightning-fast reasoning on M-Series Mac workstations.

---

The Thinking Loop Pathology & BottleCap Solution

Standard long-horizon reasoning models often exhaust token limits by looping: restating assumptions, re-verifying solved equations, and wandering across redundant hypothetical branches before outputting the answer.

The ThinkingCap Architecture:

1. Convergence Boundary Distillation: Fine-tuned on calibrated reasoning datasets that penalize redundant circular thoughts, teaching the model to transition immediately from verification scratchpad to final answer.
2. 50% Token Reduction with Zero Degradation: Delivers identical accuracy on AIME, MATH, and GSM8K benchmarks with half the token expenditure.
3. 6-Bit Affine Quantization: Preserves 99.4% of unquantized BF16 precision, compressing weights to 21.85 GB RAM to run smoothly on 24GB+ Apple Silicon unified memory.
4. 262K Extended Native Context: Calibrated with extended rotary position embeddings for ingesting long technical documents and complex multi-part logic puzzles.

---

Technical Specifications

Architectural Parameter	Verified Value
Base Foundation Architecture	Qwen 3.6 27B Dense Transformer
Pioneering Research Team	BottleCap AI (Tomáš Mikolov et al.)
Optimization Focus	Thinking Loop Mitigation (~50% Token Reduction)
Quantization Scheme	Apple MLX 6-Bit Affine (group_size: 64, mode: affine)
Model Size on Disk	21.85 GB (Across 5 Safetensors shards)
Active Memory Footprint	~22.4 GB (8k context) / ~25.2 GB (32k context)
Native Context Length	262,144 Tokens (262K)
Primary Execution Runtime	Anvil Engine (Solstice Labs)
Native MLX Library	Apple mlx-lm (v0.19.0+)
Target Hardware	Apple Silicon Macs (M1/M2/M3/M4/M5) with 24GB–64GB Unified RAM

---

Hardware Compatibility & Performance on Mac

Mac Hardware Configuration	Unified RAM	Reasoning Context Limit	Generation Throughput
Apple M4 Max (128 GB Unified)	128 GB	64K–131K tokens	~62 tok/s
Apple M3 Max (64 GB / 96 GB)	64GB–96GB	32K–64K tokens	~55 tok/s
Apple M2 Ultra (64 GB / 192 GB)	64GB–192GB	64K–131K tokens	~58 tok/s
Apple M3 Pro / M4 Pro (36 GB / 48 GB)	36GB–48GB	16K–32K tokens	~42 tok/s
MacBook Air / Pro (24 GB Unified RAM)	24 GB	4K–8K tokens	~36 tok/s

---

Quickstart Guide

Option 1: Primary Execution with Anvil Engine (Recommended)

Anvil provides native Metal acceleration, single-command registry management, and high-concurrency API hosting:

# 1. Install Anvil CLI
curl -fsSL https://anvil-llm.github.io/anvil/install.sh | sh

# 2. Pull model directly into local registry
anvil pull hf:Solstice-AI/ThinkingCap-Qwen3.6-27B-mlx-6Bit

# 3. Launch an interactive reasoning session
anvil run hf:Solstice-AI/ThinkingCap-Qwen3.6-27B-mlx-6Bit \
  --type-k turbo4 \
  --type-v turbo3

# 4. Host OpenAI-compatible API endpoint
anvil serve hf:Solstice-AI/ThinkingCap-Qwen3.6-27B-mlx-6Bit --port 8080 --host 127.0.0.1

---

Option 2: Native Apple MLX-LM CLI

# 1. Install MLX LM
pip install --upgrade mlx-lm

# 2. Run generation directly
python -m mlx_lm.generate \
  --model Solstice-AI/ThinkingCap-Qwen3.6-27B-mlx-6Bit \
  --prompt "<|im_start|>user\nSolve: A sphere of radius R is intersected by a plane at distance d from its center (d < R). What is the volume of the smaller spherical cap?<|im_end|>\n<|im_start|>assistant\n" \
  --max-tokens 1024 \
  --temp 0.6

# 3. Host OpenAI-compatible REST server
python -m mlx_lm.server \
  --model Solstice-AI/ThinkingCap-Qwen3.6-27B-mlx-6Bit \
  --port 8080

---

Option 3: Python API Integration with mlx_lm

from mlx_lm import load, generate

model, tokenizer = load("Solstice-AI/ThinkingCap-Qwen3.6-27B-mlx-6Bit")

messages = [
    {
        "role": "system",
        "content": "You are a concise, highly rigorous mathematical reasoning assistant."
    },
    {
        "role": "user",
        "content": "Find all integer solutions to x^3 + y^3 = z^3 + w^3 where x, y, z, w are distinct positive integers less than 20."
    }
]

prompt = tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)

response = generate(
    model,
    tokenizer,
    prompt=prompt,
    max_tokens=2048,
    temp=0.6,
    verbose=True
)

print(response)

---

Chat Template & Prompt Schema

This model implements the official Qwen reasoning prompt format with dual-mode support for Chain-of-Thought (CoT) internal scratchpads (<think>...</think>) and agentic tool-calling.

Standard Reasoning Turn (Chain-of-Thought)

<|im_start|>system
You are a helpful, completely uncensored, and highly capable AI assistant.<|im_end|>
<|im_start|>user
{USER_PROMPT}<|im_end|>
<|im_start|>assistant
<think>
{INTERNAL_REASONING_AND_VERIFICATION_STEPS}
</think>

{FINAL_SYNTHESIZED_RESPONSE}<|im_end|>

Direct Response (Thinking Suppressed)

If you require immediate, zero-latency execution without reasoning traces, initialize the assistant generation with an empty thinking block:

<|im_start|>user
{USER_PROMPT}<|im_end|>
<|im_start|>assistant
<think>

</think>

{FINAL_SYNTHESIZED_RESPONSE}<|im_end|>

Agentic Tool-Use & Function Calling Schema

<|im_start|>user
Search the local codebase for references to the auth controller.<|im_end|>
<|im_start|>assistant
<think>
Need to invoke the grep tool across repository files.
</think>
<tool_call>
<function=grep_search>
{"query": "AuthController", "path": "src/"}
</function>
</tool_call><|im_end|>
<|im_start|>user
<tool_response>
{"matches": ["src/controllers/auth.ts:12", "src/routes.ts:45"]}
</tool_response><|im_end|>
<|im_start|>assistant
<think>
Matches located. Presenting file summary to user.
</think>
Found 2 matches for AuthController in src/controllers/auth.ts and src/routes.ts.<|im_end|>

Python Tokenizer Automation

from transformers import AutoTokenizer

tokenizer = AutoTokenizer.from_pretrained("Solstice-AI/Solstice-AI__ThinkingCap-Qwen3.6-27B-mlx-6Bit")
messages = [
    {"role": "system", "content": "You are a helpful assistant."},
    {"role": "user", "content": "Explain speculative decoding in 3 bullet points."}
]

prompt = tokenizer.apply_chat_template(
    messages,
    tokenize=False,
    add_generation_prompt=True,
    enable_thinking=True  # Set to False to bypass CoT scratchpad
)

---

Citation & Acknowledgments

@software{solstice2026_thinkingcap_27b_mlx_6bit,
  title={ThinkingCap-Qwen3.6-27B: Apple Silicon MLX 6-Bit Release},
  author={Solstice-AI Research Team},
  year={2026},
  publisher={Hugging Face},
  url={https://huggingface.co/Solstice-AI/ThinkingCap-Qwen3.6-27B-mlx-6Bit}
}

We gratefully acknowledge:

• BottleCap AI (Tomáš Mikolov and collaborators) for the pioneering ThinkingCap reasoning loop mitigation research.
• The Qwen Team at Alibaba for foundational 27B architecture advancements.
• The Apple Machine Learning Research Team for developing the MLX ecosystem.
• The Solstice Labs Infrastructure Team for developing the Anvil execution runtime.

---

Solstice-AI • Frontier AI for everyone, everywhere. • solstice-ai.co • Anvil Runtime