Compare / Gemma 4 31B vs Kanana-2 30B-A3B Thinking

Gemma 4 31B vs Kanana-2 30B-A3B Thinking

Updated November 2026

Verdict

Gemma 4 31B scores higher on the Small LM Index: 70.1 vs 50.7. On the 4 benchmarks both labs report, Gemma 4 31B wins 4 and Kanana-2 30B-A3B Thinking wins 0. Both need about 21.9 GB of memory at 4-bit quantization. Only Gemma 4 31B has a hosted API on OpenRouter ($0.16 per million tokens); the other has to be self-hosted.

Pick Gemma 4 31B if…

you want the strongest model overall
you mostly write code
you need math and reasoning
you want a hosted API rather than running it yourself
you work with long documents
you need a permissive license (Apache-2.0)
you need image input
you want the more widely used model (more quantized builds and community support)

Pick Kanana-2 30B-A3B Thinking if…

No clear advantage in the data we have.

Specifications

	Gemma 4 31B	Kanana-2 30B-A3B Thinking
Small LM Index	70.1 (likely 69–72)	50.7 (likely 46–55)
Lab	Google · United States	Kakao (Kanana) · Rest of Asia
Released	2026-04-02	2026-01-14
Parameters	30.7B (dense)	30.7B (3B active, MoE)
Memory at 4-bit	~21.9 GB	~20.2 GB
Context window	256K tokens	32K tokens
Input	text, image	text
Reasoning	Hybrid (thinking on/off)	Always thinks
License	Apache-2.0	Kanana License
API price / 1M tokens	$0.16 ($0.10 in, $0.34 out)	Self-host only
Hugging Face downloads (30 days)	8M	1.3K

Benchmarks

Higher is better. Bold marks the winner where both labs report a score. ≈ marks the index model's estimate for a score that wasn't reported, ✓ an independent measurement and * a score measured by another lab.

Benchmark	Gemma 4 31B	Kanana-2 30B-A3B Thinking
MMLU-Pro Knowledge	85.2	74.2
GPQA Diamond Knowledge	84.3	57.8
Humanity's Last Exam Reasoning	19.5	≈ 7
AIME Math	89.2	74.0
LiveCodeBench Coding	80.0	58.8
SWE-bench Verified Coding	52.0^*	≈ 32
Terminal-Bench Agentic	42.9^*	—
τ²-Bench Agentic	76.9	—
IFBench Instruction following	—	47.8
MMMU-Pro Vision	76.9	—
IFEval Instruction following	—	82.2

Scores are mostly reported by each lab with its own settings, so small gaps are noise. See the methodology. Open in the interactive comparison to add more models.

Questions

Which is better, Gemma 4 31B or Kanana-2 30B-A3B Thinking?

Gemma 4 31B scores higher on the Small LM Index: 70.1 vs 50.7. On the 4 benchmarks both labs report, Gemma 4 31B wins 4 and Kanana-2 30B-A3B Thinking wins 0.

Which is cheaper to run?

Only Gemma 4 31B has a hosted API on OpenRouter ($0.16 per million tokens); the other has to be self-hosted. Both need about 21.9 GB of memory at 4-bit quantization.

Which has the longer context window?

Gemma 4 31B has the longer context window: 256K tokens against 32K.

Can I use Gemma 4 31B and Kanana-2 30B-A3B Thinking commercially?

Gemma 4 31B is released under Apache-2.0 and Kanana-2 30B-A3B Thinking under Kanana License. Apache 2.0 and MIT allow commercial use; other licenses can add conditions, so read the license before shipping a product.