✨ Add Terminal Bench results

results/terminal-bench-2-0-qwen-3-6-36b-nvfp4-opencode.json

@@ -0,0 +1,58 @@
+{
+    "benchmark": {
+        "name": "Terminal Bench 2.0",
+        "repo": "terminal-bench/terminal-bench-2",
+        "num_tasks": 89,
+        "url": "https://www.tbench.ai/benchmarks/terminal-bench-2"
+    },
+    "harness": {
+        "name": "OpenCode",
+        "skills": [],
+        "is_oss": true,
+        "url": "https://github.com/anomalyco/opencode"
+    },
+    "model": {
+        "name": "Qwen3.6-35B-A3B-NVFP4",
+        "repo": "RedHatAI/Qwen3.6-35B-A3B-NVFP4",
+        "is_oss": true,
+        "num_params": 35,
+        "precision": "nvfp4",
+        "url": "https://huggingface.co/RedHatAI/Qwen3.6-35B-A3B-NVFP4"
+    },
+    "environment": {
+        "name": "harbor",
+        "config": {
+            "path": null,
+            "name": "terminal-bench/terminal-bench-2",
+            "version": null,
+            "ref": "sha256:c6fc2e2382c1dbae99b2d5ecd2f4f4a60c3c01e0d84642d69b4afd92e99d078b",
+            "registry_url": null,
+            "registry_path": null,
+            "overwrite": false,
+            "download_dir": null,
+            "task_names": null,
+            "exclude_task_names": null,
+            "n_tasks": null
+        },
+        "url": "https://github.com/harbor-framework/harbor"
+    },
+    "metrics": {
+        "n_tasks": 89,
+        "n_errors": 5,
+        "score": 0.303,
+        "n_input_tokens": 47607780,
+        "n_cache_tokens": 0,
+        "n_output_tokens": 1657188,
+        "n_total_tokens": 49264968,
+        "agent_time_seconds": 46467,
+        "total_time_seconds": 58507,
+        "cost_usd": 11.47,
+        "mean_input_tokens_per_task": 534918,
+        "mean_cache_tokens_per_task": 0,
+        "mean_output_tokens_per_task": 18620,
+        "mean_tokens_per_task": 553538,
+        "mean_cost_usd_per_task": 0.13,
+        "mean_total_time_seconds_per_task": 657,
+        "mean_agent_time_seconds_per_task": 522
+    }
+}

results/terminal-bench-2-0-qwen-3-6-36b-nvfp4-pi.json

@@ -0,0 +1,58 @@
+{
+    "benchmark": {
+        "name": "Terminal Bench 2.0",
+        "repo": "terminal-bench/terminal-bench-2",
+        "num_tasks": 89,
+        "url": "https://www.tbench.ai/benchmarks/terminal-bench-2"
+    },
+    "harness": {
+        "name": "Pi",
+        "skills": [],
+        "is_oss": true,
+        "url": "https://github.com/earendil-works/pi/tree/main"
+    },
+    "model": {
+        "name": "Qwen3.6-35B-A3B-NVFP4",
+        "repo": "RedHatAI/Qwen3.6-35B-A3B-NVFP4",
+        "is_oss": true,
+        "num_params": 35,
+        "precision": "nvfp4",
+        "url": "https://huggingface.co/RedHatAI/Qwen3.6-35B-A3B-NVFP4"
+    },
+    "environment": {
+        "name": "harbor",
+        "config": {
+            "path": null,
+            "name": "terminal-bench/terminal-bench-2",
+            "version": null,
+            "ref": "sha256:c6fc2e2382c1dbae99b2d5ecd2f4f4a60c3c01e0d84642d69b4afd92e99d078b",
+            "registry_url": null,
+            "registry_path": null,
+            "overwrite": false,
+            "download_dir": null,
+            "task_names": null,
+            "exclude_task_names": null,
+            "n_tasks": null
+        },
+        "url": "https://github.com/harbor-framework/harbor"
+    },
+    "metrics": {
+        "n_tasks": 89,
+        "n_errors": 5,
+        "score": 0.36,
+        "n_input_tokens": 82108716,
+        "n_cache_tokens": 0,
+        "n_output_tokens": 2056390,
+        "n_total_tokens": 84165106,
+        "agent_time_seconds": 44991,
+        "total_time_seconds": 64760,
+        "cost_usd": 11.11,
+        "mean_input_tokens_per_task": 922569,
+        "mean_cache_tokens_per_task": 0,
+        "mean_output_tokens_per_task": 23105,
+        "mean_tokens_per_task": 945675,
+        "mean_cost_usd_per_task": 0.12,
+        "mean_total_time_seconds_per_task": 727,
+        "mean_agent_time_seconds_per_task": 505
+    }
+}