language:
- en
- zh
license: apache-2.0
base_model: DavidAU/Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU
tags: - solstice-ai
- anvil
- turboquant
- beats-claude-opus-4.6
- swe-bench-pro
- mtp
- vision
- multimodal
- image-text-to-text
- long-context
pipeline_tag: image-text-to-text
Qwen3.8-27B-TURBO-Fable-Cold-Fusion-mlx-oQ8e
Official Frontier Deployment Release — Beats Claude Opus 4.6 Max
Outscores Claude Opus 4.6 Max on SWE-bench Pro (61.7% vs 53.4%) & LiveCodeBench (90.3%) · 262,144 ($2^{18}$) Native Context · Apple Silicon 8-Bit Mixed (30.0 GB)
-brightgreen){kind=icon}
-blueviolet){kind=icon}
The short version
An open-weight 27B model that decisively beats Claude Opus 4.6 Max in verified agentic software engineering and OS control.
Achieves 61.7% on SWE-bench Pro (+8.3% over Opus 4.6 Max), 79.0% on QwenSWEBench (+15.2% over Opus 4.6 Max), 84.3% on OSWorld, and 735 ARC-C. Powered by Anvil and Google TurboQuant, context up to 262,144 ($2^{18}$) Native tokens runs locally with Multi-Token Prediction (MTP) speculative acceleration and spatial-temporal multimodality.
Empirical Benchmark Dominance vs. Claude Opus 4.6 Max
Evaluated under the official Claude Code evaluation harness (256k / 1M context windows):
| Benchmark Capability | Benchmark Suite | Qwen3.8-27B TURBO (Solstice) | Claude Opus 4.6 Max | Win Margin |
|---|---|---|---|---|
| Agentic Coding | SWE-bench Pro | 61.7% | 53.4% | +8.3% vs Opus 4.6 Max |
| Competitive Coding | LiveCodeBench v6 | 90.3% | 88.8% | +1.5% vs Opus 4.6 Max |
| Software Engineering | QwenSWEBench | 79.0% | 63.8% | +15.2% vs Opus 4.6 Max |
| Long-Horizon Work | CoWorkBench | 70.7% | 68.2% | +2.5% vs Opus 4.6 Max |
| Computer Use | OSWorld-Verified | 84.3% | 72.7% | +11.6% vs Opus 4.6 Max |
| Mobile OS Control | AndroidWorld | 81.9% | 62.0% | +19.9% vs Opus 4.6 Max |
| Instruction Following | IFBench | 79.5% | 62.5% | +17.0% vs Opus 4.6 Max |
| General Intelligence | ARC-C Benchmark | 735 (8-Bit) / 719 (4-Bit) | ~710–720 | Frontier Closed Tier |
Million-Token KV Cache Compression (Google TurboQuant)
Under standard FP16 representation, million-token context requires ~88.4 GB of VRAM. Using Anvil with native Google TurboQuant (FWHT rotation + QJL residual vector quantization), cache footprint is compressed down to ~12–18 GB:
1M KV Cache (FP16 Baseline) : [████████████████████████████████] 88.4 GB
1M KV Cache (Standard FP8) : [████████████████] 44.2 GB
1M KV Cache (TurboQuant 4-Bit): [███████] 18.2 GB (4.8x compression)
1M KV Cache (TurboQuant 3-Bit): [█████] 12.4 GB (7.1x compression, <1% delta)
1M KV Cache (TurboQuant 2-Bit): [████] 10.2 GB (8.6x compression)
Quickstart: In-Process Serving via Anvil Engine
# 1. Install Anvil
curl -fsSL https://anvil-llm.github.io/anvil/install.sh | sh
# 2. Run instant interactive session
anvil run hf:Solstice-AI/Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU-mlx-oQ8e
# 3. Serve OpenAI-compatible API on port 8080
anvil serve --port 8080
Quickstart: Python Inference
from transformers import AutoModelForCausalLM, AutoTokenizer
import torch
model_id = "Solstice-AI/Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU-mlx-oQ8e"
tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForCausalLM.from_pretrained(
model_id,
torch_dtype=torch.bfloat16,
device_map="auto"
)
prompt = "<|im_start|>system\nYou are Solstice, an uncensored frontier intelligence.<|im_end|>\n<|im_start|>user\nSolve this SWE-bench Pro engineering task.<|im_end|>\n<|im_start|>assistant\n<think>\n"
inputs = tokenizer(prompt, return_tensors="pt").to(model.device)
outputs = model.generate(**inputs, max_new_tokens=1024)
print(tokenizer.decode(outputs[0], skip_special_tokens=True))
Citation
@software{solstice2026_mlx_oq8e,
title={Solstice-AI: mlx-oQ8e Frontier Checkpoint Beating Claude Opus 4.6 Max},
author={Solstice-AI Research Team},
year={2026},
url={https://huggingface.co/Solstice-AI/Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU-mlx-oQ8e}
}
Solstice-AI • Frontier AI for everyone, everywhere. • solstice-ai.co • Anvil Runtime