Compare / Kanana-2 30B-A3B Thinking vs Qwen3-Coder 30B-A3B
Kanana-2 30B-A3B Thinking vs Qwen3-Coder 30B-A3B
Updated November 2026
Verdict
Kanana-2 30B-A3B Thinking scores higher on the Small LM Index: 54.8 vs 34.9. On the 5 benchmarks both labs report, Kanana-2 30B-A3B Thinking wins 5 and Qwen3-Coder 30B-A3B wins 0. Both need about 20.4 GB of memory at 4-bit quantization. Only Qwen3-Coder 30B-A3B has a hosted API on OpenRouter ($0.12 per million tokens); the other has to be self-hosted.
Pick Kanana-2 30B-A3B Thinking if…
- you want the strongest model overall
- you mostly write code
- you need math and reasoning
- you need precise instruction following
Pick Qwen3-Coder 30B-A3B if…
- you want a hosted API rather than running it yourself
- you work with long documents
- you need a permissive license (Apache-2.0)
- you want the more widely used model (more quantized builds and community support)
Specifications
| Kanana-2 30B-A3B Thinking | Qwen3-Coder 30B-A3B | |
|---|---|---|
| Small LM Index | 54.8 (likely 52–58) | 34.9 (likely 32–38) |
| Lab | Kakao (Kanana) · Rest of Asia | Alibaba (Qwen) · China |
| Released | 2026-01-14 | 2025-07-31 |
| Parameters | 30.7B (3B active, MoE) | 30.5B (3.3B active, MoE) |
| Memory at 4-bit | ~20.2 GB | ~20.4 GB |
| Context window | 32K tokens | 262K tokens |
| Input | text | text |
| Reasoning | Always thinks | No thinking |
| License | Kanana License | Apache-2.0 |
| API price / 1M tokens | Self-host only | $0.12 ($0.070 in, $0.28 out) |
| Hugging Face downloads (30 days) | 1.4K | 1.3M |
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 | Kanana-2 30B-A3B Thinking | Qwen3-Coder 30B-A3B |
|---|---|---|
| MMLU-Pro Knowledge | 74.2 | 70.6✓ |
| GPQA Diamond Knowledge | 57.8 | 51.6✓ |
| Humanity's Last Exam Reasoning | ≈ 9 | 3.8✓ |
| AIME Math | 74.0 | 29.7✓ |
| LiveCodeBench Coding | 58.8 | 40.3✓ |
| SWE-bench Verified Coding | ≈ 50 | 51.6 |
| τ²-Bench Agentic | — | 34.5✓ |
| IFBench Instruction following | 47.8 | 32.7✓ |
| IFEval Instruction following | 82.2 | — |
| MATH-500 Math | — | 89.3✓ |
| AA-LCR Long context | — | 32.7✓ |
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, Kanana-2 30B-A3B Thinking or Qwen3-Coder 30B-A3B?
Kanana-2 30B-A3B Thinking scores higher on the Small LM Index: 54.8 vs 34.9. On the 5 benchmarks both labs report, Kanana-2 30B-A3B Thinking wins 5 and Qwen3-Coder 30B-A3B wins 0.
Which is cheaper to run?
Only Qwen3-Coder 30B-A3B has a hosted API on OpenRouter ($0.12 per million tokens); the other has to be self-hosted. Both need about 20.4 GB of memory at 4-bit quantization.
Which has the longer context window?
Qwen3-Coder 30B-A3B has the longer context window: 262K tokens against 32K.
Can I use Kanana-2 30B-A3B Thinking and Qwen3-Coder 30B-A3B commercially?
Kanana-2 30B-A3B Thinking is released under Kanana License and Qwen3-Coder 30B-A3B under Apache-2.0. Apache 2.0 and MIT allow commercial use; other licenses can add conditions, so read the license before shipping a product.
Related comparisons
- Gemma 4 31B vs Kanana-2 30B-A3B Thinking
- GLM-4.7-Flash vs Kanana-2 30B-A3B Thinking
- Kanana-2 30B-A3B Thinking vs Qwen3 30B-A3B Thinking 2507
- Muse Glimmer 30B vs Qwen3-Coder 30B-A3B
- Gemma 4 31B vs Qwen3-Coder 30B-A3B
- GLM-4.7-Flash vs Qwen3-Coder 30B-A3B
- Granite 4.2 30B vs Qwen3-Coder 30B-A3B
- Qwen3 30B-A3B Thinking 2507 vs Qwen3-Coder 30B-A3B
- North Mini Code vs Qwen3-Coder 30B-A3B