[
  {
    "benchmark": "gaia",
    "score": 33.9,
    "metric": "accuracy",
    "cost_per_instance": 0.28,
    "average_runtime": 197.0,
    "full_archive": "https://results.eval.all-hands.dev/eval-21093373009-qwen-3-cod_litellm_proxy-fireworks_ai-qwen3-coder-480b-a35b-instruct_26-01-17-12-54.tar.gz",
    "tags": [
      "gaia"
    ],
    "agent_version": "v1.8.3",
    "submission_time": "2026-01-26T15:55:58.082973+00:00"
  },
  {
    "benchmark": "swe-bench",
    "score": 62.4,
    "metric": "accuracy",
    "cost_per_instance": 1.26,
    "average_runtime": 680.0,
    "full_archive": "https://results.eval.all-hands.dev/eval-20979851181-qwen-3-cod_litellm_proxy-fireworks_ai-qwen3-coder-480b-a35b-instruct_26-01-14-09-02.tar.gz",
    "tags": [
      "swe-bench"
    ],
    "agent_version": "v1.8.3",
    "submission_time": "2026-01-26T15:55:58.082973+00:00"
  },
  {
    "benchmark": "commit0",
    "score": 0.0,
    "metric": "accuracy",
    "cost_per_instance": 0.01,
    "average_runtime": 1.0,
    "full_archive": "https://results.eval.all-hands.dev/commit0/litellm_proxy-fireworks_ai-qwen3-coder-480b-a35b-instruct/22104205466/results.tar.gz",
    "tags": [
      "commit0"
    ],
    "agent_version": "v1.11.0",
    "submission_time": "2026-02-17T15:35:19+00:00"
  },
  {
    "benchmark": "swt-bench",
    "score": 34.9,
    "metric": "accuracy",
    "cost_per_instance": 0.97,
    "average_runtime": 626.0,
    "full_archive": "https://results.eval.all-hands.dev/eval-21179579508-qwen-3-cod_litellm_proxy-fireworks_ai-qwen3-coder-480b-a35b-instruct_26-01-21-01-51.tar.gz",
    "tags": [
      "swt-bench"
    ],
    "agent_version": "v1.8.3",
    "submission_time": "2026-01-26T15:55:58.082973+00:00"
  },
  {
    "benchmark": "swe-bench-multimodal",
    "score": 23.5,
    "metric": "solveable_accuracy",
    "cost_per_instance": 2.09,
    "average_runtime": 1006.0,
    "full_archive": "https://results.eval.all-hands.dev/eval-21357041327-qwen-3-cod_litellm_proxy-fireworks_ai-qwen3-coder-480b-a35b-instruct_26-01-26-15-08.tar.gz",
    "tags": [
      "swe-bench-multimodal"
    ],
    "component_scores": {
      "solveable_accuracy": 23.5,
      "unsolveable_accuracy": 5.9,
      "combined_accuracy": 17.6
    },
    "agent_version": "v1.8.3",
    "submission_time": "2026-01-26T15:55:58.082973+00:00"
  }
]