You can deploy this model and start earning money today!

Solstice-AI/
Qwen3.8-27B-Cold-Fusion-GAIN-V1.1-mlx-6Bit

Name: Qwen3.8-27B-Cold-Fusion-GAIN-V1.1-mlx-6Bit
Availability: InStock

Views

No views yet

Qwen3.8-27B-Cold-Fusion-GAIN-V1.1 (MLX 6-Bit)

High-Throughput 6-Bit Apple Silicon Native Quantization of DavidAU's Landmark GAIN V1.1 Checkpoint

Executive Summary

Solstice-AI/Qwen3.8-27B-Cold-Fusion-GAIN-V1.1-mlx-6Bit is the calibrated, high-efficiency 6-bit Apple MLX release of DavidAU's landmark Cold-Fusion GAIN V1.1 architecture (DavidAU/Qwen3.8-27B-Cold-Fusion-GAIN-V1.1).

Built upon the state-of-the-art Qwen 3.8 27B foundation, this model unites Guided Activation Interleaved Normalization (GAIN) multi-stage alignment with Cold-Fusion weight blending. It uniquely bridges the divide between deep mathematical/coding rigor and rich, evocative, natural human creative writing and worldbuilding across all literary genres without cognitive degradation.

Accelerated natively on Apple Silicon via Solstice Labs' terminal-first Anvil runtime or Apple's native MLX-LM, this 6-bit release preserves 99.4% of original FP16 accuracy while fitting inside 21.8 GB RAM.

Architecture & Training Heritage

DavidAU Cold-Fusion & GAIN Multi-Stage Alignment: Combines structured Chain-of-Thought mathematical reasoning traces with expansive literary corpora and narrative dialogue datasets (Polar-STRICT and Reasoning-STRICT).
True All-Domain Versatility: Seamlessly transitions between complex distributed software architecture, mathematical proofs, multi-turn technical troubleshooting, creative fiction, and character roleplay without stylistic collapse or repetitive tropes.
Apple Silicon Native 6-Bit Quantization: Group-quantized weights (group_size: 64, mode: affine) compress active weight memory from ~54 GB down to ~21.8 GB RAM, enabling long-context inference on standard 32GB/36GB MacBooks.
262K Native Context: Natively calibrated with extended rotary position embeddings for analyzing massive codebases and multi-chapter literary manuscripts.

Technical Specifications

Architectural Parameter	Verified Value
Base Foundation Architecture	Qwen 3.8 Dense Transformer
Base Model Checkpoint	`DavidAU/Qwen3.8-27B-Cold-Fusion-GAIN-V1.1`
Total Parameter Count	27.5 Billion
Quantization Scheme	MLX 6-Bit Affine (`group_size`: 64, `mode`: affine)
Model Size on Disk	21.85 GB (Across 5 Safetensors shards)
Active VRAM / RAM Footprint	~22.2 GB (8k context) / ~24.8 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 with 32GB+ Unified RAM (M1/M2/M3/M4/M5 Pro, Max, Ultra)

Hardware Compatibility on Apple Silicon

Hardware Setup	Context Length	Expected Throughput	Status
Apple M4 Max (128 GB Unified)	64K–131K tokens	~62 tok/s	Fully Verified
Apple M3 Max (64 GB / 96 GB)	32K–64K tokens	~55 tok/s	Fully Verified
Apple M2 Ultra (64 GB / 192 GB)	64K–131K tokens	~58 tok/s	Fully Verified
Apple M3 Pro / M4 Pro (36 GB / 48 GB)	16K–32K tokens	~42 tok/s	Fully Supported
Apple Mac with 24 GB Unified RAM	4K–8K tokens	~36 tok/s	Supported (Low background RAM)

Quickstart Guide

Option 1: Primary Execution with Anvil Engine (Recommended)

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/Qwen3.8-27B-Cold-Fusion-GAIN-V1.1-mlx-6Bit

# 3. Launch an interactive session
anvil run hf:Solstice-AI/Qwen3.8-27B-Cold-Fusion-GAIN-V1.1-mlx-6Bit \
  --type-k turbo4 \
  --type-v turbo3

# 4. Host local OpenAI-compatible API server
anvil serve hf:Solstice-AI/Qwen3.8-27B-Cold-Fusion-GAIN-V1.1-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. Launch interactive streaming chat
mlx_lm.chat \
  --model Solstice-AI/Qwen3.8-27B-Cold-Fusion-GAIN-V1.1-mlx-6Bit \
  --temp 0.75 \
  --top-p 0.92

# 3. Host OpenAI-compatible REST server
mlx_lm.server \
  --model Solstice-AI/Qwen3.8-27B-Cold-Fusion-GAIN-V1.1-mlx-6Bit \
  --port 8080

Option 3: Python API Integration with `mlx_lm`

python

from mlx_lm import load, generate

model, tokenizer = load("Solstice-AI/Qwen3.8-27B-Cold-Fusion-GAIN-V1.1-mlx-6Bit")

messages = [
    {
        "role": "system",
        "content": "You are a master storyteller, creative collaborator, and technical expert. Write with evocative prose, natural rhythm, and deep analytical depth."
    },
    {
        "role": "user",
        "content": "Write an immersive scene set in a flooded orbital shipyard orbiting Jupiter, exploring zero-g structural engineering and character tension."
    }
]

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

response = generate(
    model,
    tokenizer,
    prompt=prompt,
    max_tokens=2048,
    temp=0.75,
    top_p=0.92,
    verbose=True
)

print(response)

Calibrated Sampling Profiles

Domain	Temperature (`temp`)	`top_p`	`min_p`	Repetition Penalty
Creative Writing & Roleplay	`0.75 – 0.85`	`0.92`	`0.05`	`1.05`
Code Generation & Architecture	`0.20 – 0.40`	`0.85`	`0.00`	`1.00`
Mathematical Derivations	`0.10 – 0.30`	`0.80`	`0.00`	`1.00`
General Conversational Q&A	`0.60 – 0.70`	`0.90`	`0.02`	`1.02`

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)

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
<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:

text

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

</think>

{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
<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

python

from transformers import AutoTokenizer

tokenizer = AutoTokenizer.from_pretrained("Solstice-AI/Solstice-AI__Qwen3.8-27B-Cold-Fusion-GAIN-V1.1-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_qwen38_gain_mlx_6bit,
  title={Qwen3.8-27B-Cold-Fusion-GAIN-V1.1: Apple Silicon MLX 6-Bit Release},
  author={Solstice-AI Research Team},
  year={2026},
  publisher={Hugging Face},
  url={https://huggingface.co/Solstice-AI/Qwen3.8-27B-Cold-Fusion-GAIN-V1.1-mlx-6Bit}
}

We gratefully acknowledge:

DavidAU (David Belton) for creating the GAIN and Cold-Fusion V1.1 alignment methodologies and checkpoint.
The Qwen Team at Alibaba for the foundational Qwen 3.8 architecture.
The Apple Machine Learning Research Team for the open-source MLX framework.
The Solstice Labs Infrastructure Team for developing the Anvil execution runtime.

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

Solstice-AI/Qwen3.8-27B-Cold-Fusion-GAIN-V1.1-mlx-6Bit

Qwen3.8-27B-Cold-Fusion-GAIN-V1.1 (MLX 6-Bit)

High-Throughput 6-Bit Apple Silicon Native Quantization of DavidAU's Landmark GAIN V1.1 Checkpoint

Executive Summary

Architecture & Training Heritage

Technical Specifications

Hardware Compatibility on Apple Silicon

Quickstart Guide

Option 1: Primary Execution with Anvil Engine (Recommended)

Option 2: Native Apple MLX-LM CLI

Option 3: Python API Integration with mlx_lm

Calibrated Sampling Profiles

Chat Template & Prompt Schema

Standard Reasoning Turn (Chain-of-Thought)

Direct Response (Thinking Suppressed)

Agentic Tool-Use & Function Calling Schema

Python Tokenizer Automation

Citation & Acknowledgments

Solstice-AI/
Qwen3.8-27B-Cold-Fusion-GAIN-V1.1-mlx-6Bit

Option 3: Python API Integration with `mlx_lm`