You can deploy this model and start earning money today!

Viesar/
gemma-3-4b-opus-reasoning-distill

Name: gemma-3-4b-opus-reasoning-distill
Availability: InStock

text-generation-inference

endpoints_compatible

Views

No views yet

Gemma 3 4B — Opus 4.6 Reasoning Distillation (Merged FP16)

QLoRA fine-tune of google/gemma-3-4b-it, with adapters merged into a single FP16 model.

This is the transformers format version, suitable for further fine-tuning, evaluation, or running with full-precision inference.

For local inference, use the GGUF version (Q4_K_M, ~2.5 GB, multimodal-ready): Viesar/gemma-3-4b-opus-reasoning-distill-GGUF

Training Summary

Base: unsloth/gemma-3-4b-it-unsloth-bnb-4bit
Method: QLoRA (r=16, alpha=16)
Dataset: Crownelius/Opus-4.6-Reasoning-3300x, filtered to ~1,900 examples
Hardware: NVIDIA RTX 3050 8GB
Duration: 50 minutes
Final loss: 1.01

Benchmark Results

Both models evaluated under identical conditions (4-bit quant, bf16, batch=1, deterministic generation) using lm-eval-harness 0.4.11.

MATH-500 (full test set, 500 problems, 4-shot)

Metric	Base Gemma 3 4B	This Model	Δ
`exact_match`	29.6% ± 2.0%	24.6% ± 1.9%	−5.0pp
`math_verify`	36.4% ± 2.2%	29.6% ± 2.0%	−6.8pp

GSM8K (300-problem subset, 5-shot)

Metric	Base Gemma 3 4B	This Model	Δ
`exact_match` (strict-match)	68.7% ± 2.7%	53.7% ± 2.9%	−15.0pp
`exact_match` (flexible)	69.0% ± 2.7%	55.0% ± 2.9%	−14.0pp

The fine-tune underperforms the base on both benchmarks. This is reported transparently. Two effects likely contribute:

Format mismatch — both benchmarks use few-shot completion prompts (Problem:...Solution: and Question:...Answer:) that don't match the chat-template <think>...</think> style the model was trained on. The model's trained instinct can fail the regex answer extractor even when reasoning is correct.
Capability narrowing — training on ~1,900 narrow examples (94% math) likely traded some general flexibility for the targeted reasoning style.

The qualitative outputs (visible on the GGUF repo) show the trained reasoning style is intact. The model is more useful for chat-style reasoning tasks than these completion-style benchmarks suggest.

Full analysis: see the GGUF README.

Performance (RTX 3050 8GB, 4-bit)

Metric	Base	Fine-tune
Per-problem time (MATH-500)	16.5 s	16.6 s
Per-problem time (GSM8K)	11.0 s	~11 s
Peak VRAM	~6.2 GB	~6.2 GB

Inference cost is identical to the base.

Usage

python

from transformers import AutoModelForCausalLM, AutoTokenizer

model = AutoModelForCausalLM.from_pretrained("Viesar/gemma-3-4b-opus-reasoning-distill")
tokenizer = AutoTokenizer.from_pretrained("Viesar/gemma-3-4b-opus-reasoning-distill")

messages = [{"role": "user", "content": "Solve: 2x^2 + 5x - 12 = 0"}]
prompt = tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
inputs = tokenizer(prompt, return_tensors="pt")
outputs = model.generate(**inputs, max_new_tokens=512)
print(tokenizer.decode(outputs[0]))

License

Gemma License