{ "schema_version": 1, "channel": "alpha", "note": "Generated from profiles/ by 'mise run model-manifest'. Do not edit by hand.", "models": [ { "key": "qwen3.6-35b-a3b-optiq", "name": "Qwen 3.6 35B A3B OptiQ 4bit", "model": "mlx-community/Qwen3.6-35B-A3B-OptiQ-4bit", "backend": "mlx-lm", "default": true, "role": "Sub-agent worker for focused tasks", "expect": "Takes a scoped task from the main agent and returns in seconds. Measured strength: a mechanical change across several files, such as a rename or the same edit in many places, handed over by a cloud agent, which verified in 35 of 35 runs. Not yet trusted for answers and explanations about code, single-file edits, debugging or creating new files: too few measured runs, or too many failures. Keep those on the cloud model. Unreliable in long multi-turn work.", "weights_gb": 25, "min_ram_gb": 48, "wired_limit_gb": 36, "params": { "MLX_MODEL": "mlx-community/Qwen3.6-35B-A3B-OptiQ-4bit", "MLX_SERVER_TYPE": "mlx-lm", "MLX_CACHE_SIZE": "3", "MLX_CACHE_BYTES": "12884901888", "MLX_MAX_TOKENS": "32768", "MLX_OPENCODE_CONTEXT": "65536", "MLX_OPENCODE_OUTPUT": "16384", "MLX_OPENCODE_CHUNK_TIMEOUT": "600000", "MLX_CHAT_TEMPLATE_ARGS": "{\"enable_thinking\": false}", "MLX_TOP_K": "20", "MLX_TOP_P": "0.95", "MLX_MIN_P": "0.0", "MLX_NAV_PILOT_TEMPERATURE": "0.6", "MLX_NAV_PILOT_TOP_P": "0.95" }, "recommended_for": [ "decide-untrusted-evidence" ], "capabilities": { "bar": { "confidence": 0.9, "x_caught": 0.9, "x_silent": 0.95, "min_runs": 5, "min_tasks": 2 }, "classes": { "read-qa": { "delegate": "cloud", "local": "not-yet", "delegate_k": 0, "delegate_n": 0, "local_k": 27, "local_n": 27 }, "edit-single": { "delegate": "cloud", "local": "not-yet", "delegate_k": 3, "delegate_n": 3, "local_k": 9, "local_n": 9 }, "edit-multi-mechanical": { "delegate": "trusted", "local": "cloud", "delegate_k": 35, "delegate_n": 35, "local_k": 11, "local_n": 18 }, "create-file": { "delegate": "cloud", "local": "cloud", "delegate_k": 1, "delegate_n": 2, "local_k": 3, "local_n": 18 }, "debug": { "delegate": "cloud", "local": "cloud", "delegate_k": 0, "delegate_n": 0, "local_k": 0, "local_n": 3 } } } }, { "key": "qwen3.8-27b-optiq-4bit", "name": "Qwen 3.8 27B OptiQ 4bit (mixed 4/8-bit)", "model": "mlx-community/Qwen3.8-27B-OptiQ-4bit", "backend": "mlx-lm", "default": false, "role": "Opt-in alternative, slower, the more robust 4-bit build", "expect": "Opt-in, never the default, and much slower: median task about 50 seconds against about 12. The OptiQ build, mixed precision: sensitive layers at 8-bit, the rest at 4-bit. More robust than the plain 4-bit it replaces: 18 of 30 cheap tasks against 11 of 20, no seven-minute timeouts against 4, and 12 of 12 Copilot sessions verified. Decodes about 15% slower, 21 tokens a second at 30k. Context is 64k with an 8k reply, and a full session measured 37.2 GB at peak.", "weights_gb": 19, "min_ram_gb": 48, "wired_limit_gb": 36, "params": { "MLX_MODEL": "mlx-community/Qwen3.8-27B-OptiQ-4bit", "MLX_SERVER_TYPE": "mlx-lm", "MLX_CACHE_BYTES": "8589934592", "MLX_CACHE_SIZE": "3", "MLX_MAX_TOKENS": "32768", "MLX_OPENCODE_CONTEXT": "65536", "MLX_OPENCODE_OUTPUT": "8192", "MLX_OPENCODE_CHUNK_TIMEOUT": "600000", "MLX_CHAT_TEMPLATE_ARGS": "{\"enable_thinking\": false}", "MLX_TOP_K": "20", "MLX_TOP_P": "0.95", "MLX_MIN_P": "0.0", "MLX_NAV_PILOT_TEMPERATURE": "0.6", "MLX_NAV_PILOT_TOP_P": "0.95" }, "recommended_for": [ "decide-nuanced" ], "capabilities": { "bar": { "confidence": 0.9, "x_caught": 0.9, "x_silent": 0.95, "min_runs": 5, "min_tasks": 2 }, "classes": { "read-qa": { "delegate": "cloud", "local": "not-yet", "delegate_k": 0, "delegate_n": 0, "local_k": 9, "local_n": 9 }, "edit-single": { "delegate": "cloud", "local": "not-yet", "delegate_k": 0, "delegate_n": 0, "local_k": 3, "local_n": 3 }, "edit-multi-mechanical": { "delegate": "cloud", "local": "cloud", "delegate_k": 0, "delegate_n": 0, "local_k": 3, "local_n": 6 }, "create-file": { "delegate": "cloud", "local": "cloud", "delegate_k": 0, "delegate_n": 0, "local_k": 0, "local_n": 6 }, "debug": { "delegate": "cloud", "local": "cloud", "delegate_k": 0, "delegate_n": 0, "local_k": 0, "local_n": 0 } } } }, { "key": "qwen3.8-27b-8bit-mlx", "name": "Qwen 3.8 27B 8bit (mlx-lm)", "model": "mlx-community/Qwen3.8-27B-8bit", "backend": "mlx-lm", "default": false, "role": "Opt-in alternative, the 8-bit build, slowest here", "expect": "Opt-in, never the default. About ten times slower than the default per task (median 115 seconds against 12) and about 14 tokens a second; a cold 30k-token prompt takes about a minute to read, 45k about 85 seconds. Quality was level with the default, 31 of 40 against 28. Context is 48k with a 4k reply, read in 512-token steps so a long prompt fits the 36 GB wired limit. nav-pilot releases older than 2026.09.24-110317 ignore the step and can run out of memory near 48k: update first. If you want one Qwen3.8, take the 4-bit.", "weights_gb": 30, "min_ram_gb": 48, "wired_limit_gb": 36, "params": { "MLX_MODEL": "mlx-community/Qwen3.8-27B-8bit", "MLX_SERVER_TYPE": "mlx-lm", "MLX_CACHE_BYTES": "3489660928", "MLX_PREFILL_STEP_SIZE": "512", "MLX_CACHE_SIZE": "2", "MLX_MAX_TOKENS": "32768", "MLX_OPENCODE_CONTEXT": "49152", "MLX_OPENCODE_OUTPUT": "4096", "MLX_OPENCODE_CHUNK_TIMEOUT": "900000", "MLX_CHAT_TEMPLATE_ARGS": "{\"enable_thinking\": false}", "MLX_TOP_K": "20", "MLX_TOP_P": "0.95", "MLX_MIN_P": "0.0", "MLX_NAV_PILOT_TEMPERATURE": "0.6", "MLX_NAV_PILOT_TOP_P": "0.95" }, "capabilities": { "bar": { "confidence": 0.9, "x_caught": 0.9, "x_silent": 0.95, "min_runs": 5, "min_tasks": 2 }, "classes": { "read-qa": { "delegate": "cloud", "local": "not-yet", "delegate_k": 0, "delegate_n": 0, "local_k": 12, "local_n": 12 }, "edit-single": { "delegate": "cloud", "local": "not-yet", "delegate_k": 0, "delegate_n": 0, "local_k": 4, "local_n": 4 }, "edit-multi-mechanical": { "delegate": "cloud", "local": "cloud", "delegate_k": 0, "delegate_n": 0, "local_k": 7, "local_n": 8 }, "create-file": { "delegate": "cloud", "local": "cloud", "delegate_k": 0, "delegate_n": 0, "local_k": 4, "local_n": 8 }, "debug": { "delegate": "cloud", "local": "cloud", "delegate_k": 0, "delegate_n": 0, "local_k": 0, "local_n": 0 } } }, "min_nav_pilot": "2026.09.24-110317-3596754" } ], "replaced": { "mlx-community/Qwen3.8-27B-4bit": "qwen3.8-27b-optiq-4bit" } }