Qwen3-Coder Original Chart - Spider2: leaderboard
Original Qwen3-Coder-30B-A3B-Instruct model-card chart, native six-model comparison. The exact same chart is mirrored by the FP8 card, but its headers identify unqualified 30B-A3B-Instruct and 480B-A35B-Instruct checkpoints; no FP8 or BF16 evaluation precision is stated. The decorative Qwen3-Coder-Flash title does not replace the explicit checkpoint column labels. SWE-bench Verified OpenHands 500-turn, OpenHands 100-turn and private-scaffolding rows stay separate. Private scaffolding does not establish a common implementation across models; those implementations are unreported. For other outcomes the chart supplies no common harness commit, effort, sampling, token budget, benchmark subset/revision or repeat count; none is inferred. BFCL explicitly uses v3. Native chart values are percentage-scale task results; missing cells remain absent. Existing Aider and LLMStats rows have different or unbound protocols/cohorts; matching numbers alone do not establish interchangeable runs or authority to overwrite them. Native outcome: Spider2. Metric: Native task result (%).
Metric: Native task result (%). Source: huggingface.co. Saturation forecast: Around June 2027. 6 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Claude Sonnet 4 | 31.1 | #203 |
| 2 | Qwen 3 Coder 480B A35B Instruct | 31.1 | #308 |
| 3 | GPT-4.1 | 25.6 | #250 |
| 4 | Kimi K2 | 25.2 | #245 |
| 5 | Qwen 3 Coder 30B A3B Instruct | 21.4 | #391 |
| 6 | DeepSeek V3 (0324) | 17.7 | #347 |
No result here: #3 Claude Opus 5.5, #5 GPT-6 Astra, #6 GPT-6.1 Sol.
Interactive version: theaggregate.ai/benchmark?slug=qwen3-coder-original-chart-spider2 · How It Works · Data refreshed daily, snapshot 2026-10-27.