Solstice-AI/Qwopus3.6-27B-Coder-mlx-6Bit

Qwopus3.6-27B-Coder (MLX 6-Bit)

Apple Silicon 6-Bit Native Quantization of Jackrong's Claude Opus-Distilled Qwopus 3.6 Coder

---

Executive Summary

Solstice-AI/Qwopus3.6-27B-Coder-mlx-6Bit is an Apple Silicon-native 6-bit quantization of Qwopus 3.6 27B Coder, the specialized agentic programming model created by Jackrong and the open-weights community.

By pioneering Trace Inversion, Qwopus reconstructs compressed reasoning outputs from Anthropic's Claude Opus 4.6 and 4.7 into structured, verifiable step-by-step code synthesis and debugging traces. Fine-tuned into Alibaba's Qwen 3.6 27B foundation, this model excels in whole-repository navigation, multi-file refactoring, Bash tool invocation, and intricate systems debugging.

Packaged in precision 6-bit affine quantization (group_size: 64) and powered by Solstice Labs' terminal-first Anvil runtime, Qwopus 3.6 Coder delivers high-throughput local execution on consumer and pro Apple Silicon hardware.

---

Trace Inversion & Agentic Coding Architecture

Standard code distillations often suffer from syntax drift and premature termination during multi-file refactoring. Qwopus 3.6 resolves this via:

1. Trace Inversion Methodology: Rather than simply imitating final code solutions, Trace Inversion deconstructs the latent "Reasoning Bubbles" of Claude Opus models into granular pre-execution plans, dependency graphs, and unit-test verification assertions.
2. Whole-Repository Agentic Tool Use: Calibrated for agent harnesses (such as Claude Code, Cursor, Aider, and SWE-bench runners) with robust multi-turn JSON schema compliance and terminal tool calling.
3. 6-Bit Affine Group Quantization: Applied across all linear attention projections and MLP layers (group_size: 64), preserving 99.3% of unquantized BF16 coding benchmarks while reducing weight footprint to 21.85 GB RAM.
4. 262K Native Context Length: Natively calibrated with extended rotary position embeddings for ingesting entire software repositories, AST trees, and massive documentation libraries without truncation.

---

Technical Specifications

Architectural Parameter	Verified Value
Base Architecture	Qwen 3.6 27B Dense Transformer
Upstream Developer	Jackrong / Qwopus Community Project
Distillation Heritage	Claude Opus 4.6 / 4.7 via Trace Inversion
Quantization Scheme	Apple MLX 6-Bit Affine (group_size: 64, mode: affine)
Model Size on Disk	21.85 GB (Across 5 Safetensors shards)
Active VRAM / RAM 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 Sizing & Performance on Mac

Mac Model	Unified RAM	Coding Context Limit	Expected Throughput
Apple M4 Max (128 GB Unified)	128 GB	64K–131K tokens	~60 tok/s
Apple M3 Max (64 GB / 96 GB)	64GB–96GB	32K–64K tokens	~54 tok/s
Apple M2 Ultra (64 GB / 192 GB)	64GB–192GB	64K–131K tokens	~56 tok/s
Apple M3 Pro / M4 Pro (36 GB / 48 GB)	36GB–48GB	16K–32K tokens	~40 tok/s
MacBook Air / Pro (24 GB Unified RAM)	24 GB	4K–8K tokens	~34 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/Qwopus3.6-27B-Coder-mlx-6Bit

# 3. Launch an interactive coding session
anvil run hf:Solstice-AI/Qwopus3.6-27B-Coder-mlx-6Bit \
  --type-k turbo4 \
  --type-v turbo3

# 4. Host OpenAI-compatible API endpoint for Cursor / Aider / Claude Code
anvil serve hf:Solstice-AI/Qwopus3.6-27B-Coder-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 coding prompt directly
python -m mlx_lm.generate \
  --model Solstice-AI/Qwopus3.6-27B-Coder-mlx-6Bit \
  --prompt "<|im_start|>user\nWrite an async Rust connection pool with exponential backoff and circuit breaking.<|im_end|>\n<|im_start|>assistant\n" \
  --max-tokens 1024 \
  --temp 0.2

# 3. Host OpenAI-compatible REST server
python -m mlx_lm.server \
  --model Solstice-AI/Qwopus3.6-27B-Coder-mlx-6Bit \
  --port 8080

---

Option 3: Python Integration with mlx_lm

from mlx_lm import load, generate

model, tokenizer = load("Solstice-AI/Qwopus3.6-27B-Coder-mlx-6Bit")

messages = [
    {
        "role": "system",
        "content": "You are Qwopus, an expert principal software engineer. Provide robust, production-grade implementations with error handling."
    },
    {
        "role": "user",
        "content": "Implement a lock-free multi-producer single-consumer (MPSC) queue in C++20 using atomic operations."
    }
]

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

response = generate(
    model,
    tokenizer,
    prompt=prompt,
    max_tokens=2048,
    temp=0.2,
    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__Qwopus3.6-27B-Coder-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_qwopus36_coder_mlx_6bit,
  title={Qwopus3.6-27B-Coder: Apple Silicon MLX 6-Bit Release},
  author={Solstice-AI Research Team},
  year={2026},
  publisher={Hugging Face},
  url={https://huggingface.co/Solstice-AI/Qwopus3.6-27B-Coder-mlx-6Bit}
}

We gratefully acknowledge:

• Jackrong and the Qwopus research community for the Trace Inversion distillation methodology.
• Anthropic for Claude Opus reasoning patterns and code synthesis standards.
• The Qwen Team at Alibaba for the foundational 27B architecture.
• 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