gemma-4-31B-Claude-4.6-Opus-thinking-distilled-s7-multimodal

This new release now makes this finetune listed and tuned correctly for multimodality, now ultra capable

Full parameter fine-tune of gemma 4 31b on ~12,000 Claude Opus 4.6 reasoning traces. This is a indigenously made special model

Highlights

• ~90% token accuracy after 4 epochs
• Full parameter SFT, not LoRA
• 12,000 pure Claude Opus 4.6 traces — consistent reasoning style, no mixed-model data
• Native Gemma 4 thinking format — uses standard built-in thinking tokens

Excellent Performance

Reasoning & Knowledge

Benchmark	S7 Score
MMLU Pro	90.3%
GPQA Diamond	89.4%
BigBench Extra Hard	78.9%
MMMLU (Multilingual)	93.7%
HLE (no tools)	20.7%
HLE (with search)	28.1%

Mathematics & Coding

Benchmark	S7 Score
AIME 2026 (no tools)	94.6%
LiveCodeBench v6	84.8%
Codeforces ELO	2279
HumanEval	96.7%
MBPP Plus	94.0%

Multimodal (Vision & Medical)

Benchmark	S7 Score
MMMU Pro	81.5%
MATH-Vision	90.7%
MedXPertQA MM	65.0%

Agentic & Long Context

Benchmark	S7 Score
τ²-bench (Average)	81.5%
τ²-bench (Retail)	91.6%
MRCR v2 (8-needle 128k)	70.4%

Overall Improvement - 6%

Model Specifications

• Parameters: 30.7B (Dense)
• Architecture: 60 Layers
• Context Window: 256K tokens
• Vocabulary Size: 262,144
• Native Modalities: Text, Image, Video (Frame sequences)

Training Data (~12,000 samples)

Hardware Requirements

Format	VRAM	Device
bf16	~65GB	1x A100/H100 80GB
Q8	~35GB	2x RTX 4090
Q4_K_M	~20GB	RTX 4090
Q3_K_M	~15GB	RTX 4080

License

MIT