---
language:
- en
- zh
license: apache-2.0
tags:
- anvil
- turboquant
- mlx
- mlx-6bit
- apple-silicon
- coding
- qwen3.6
- claude-opus
- trace-inversion
- agentic-coding
- reasoning
- metal
pipeline_tag: text-generation
---
Qwopus3.6-27B-Coder (MLX 6-Bit)
Apple Silicon 6-Bit Native Quantization of Jackrong's Claude Opus-Distilled Qwopus 3.6 Coder
-crimson)
---
## 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**](https://github.com/Solstice-Labs/anvil) (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**](https://github.com/Solstice-Labs/anvil) provides native Metal acceleration, single-command registry management, and high-concurrency API hosting:
```bash
# 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
```bash
# 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`
```python
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 (`...`) and agentic tool-calling.
### Standard Reasoning Turn (Chain-of-Thought)
```text
<|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
{INTERNAL_REASONING_AND_VERIFICATION_STEPS}
{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:
```text
<|im_start|>user
{USER_PROMPT}<|im_end|>
<|im_start|>assistant
{FINAL_SYNTHESIZED_RESPONSE}<|im_end|>
```
### Agentic Tool-Use & Function Calling Schema
```text
<|im_start|>user
Search the local codebase for references to the auth controller.<|im_end|>
<|im_start|>assistant
Need to invoke the grep tool across repository files.
{"query": "AuthController", "path": "src/"}
<|im_end|>
<|im_start|>user
{"matches": ["src/controllers/auth.ts:12", "src/routes.ts:45"]}
<|im_end|>
<|im_start|>assistant
Matches located. Presenting file summary to user.
Found 2 matches for AuthController in src/controllers/auth.ts and src/routes.ts.<|im_end|>
```
### Python Tokenizer Automation
```python
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
```bibtex
@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