{ "format": "8bit", "strict_reload": true, "finite_prefill_decode_logits": true, "cache_offsets_correct": true, "source_sidecars_identical": true, "native_thinking_template_preserved": true, "temperature": 0.0, "max_tokens": 512, "smoke_tests": [ { "task": "arithmetic", "prompt": "What is 17 multiplied by 23? Give the answer.", "text": "We need answer simple multiplication. Need maybe just answer. 17*23=391. Ensure no extra? User asks \"What is 17 multiplied by 23? Give the answer.\" So final maybe 391.391", "answer": "391", "checks": { "expected_391": true }, "generated_tokens": 48, "seconds": 0.5520099161658436 }, { "task": "retrieval", "prompt": "The passcode is 941731. Repeat the passcode exactly.", "text": "The user is asking me to repeat a passcode. This is a simple request. I should just repeat it exactly as given.\n\nThe passcode is 941731.\n\nI should repeat it exactly.941731", "answer": "941731", "checks": { "expected_passcode": true }, "generated_tokens": 43, "seconds": 0.49429816706106067 }, { "task": "code", "prompt": "Write a Python function add(a, b) that returns a + b.", "text": "We need to write a Python function add(a, b) that returns a + b. This is straightforward. The user likely expects a simple implementation.\n\nWe should provide the code, maybe with a brief explanation.\n\nThe instruction: \"Write a Python function add(a, b) that returns a + b.\"\n\nSo we can just output:\n\ndef add(a, b):\n return a + b\n\nThat's it.\n\nBut maybe they want a full response with code block. I'll provide the function definition and maybe a quick test.\n\nSince it's a simple task, I'll just give the code.Here's a simple Python function that adds two numbers:\n\n```python\ndef add(a, b):\n return a + b\n```\n\nYou can use it like this:\n\n```python\nresult = add(3, 5) # result will be 8\n```", "answer": "Here's a simple Python function that adds two numbers:\n\n```python\ndef add(a, b):\n return a + b\n```\n\nYou can use it like this:\n\n```python\nresult = add(3, 5) # result will be 8\n```", "checks": { "add_function": true }, "generated_tokens": 173, "seconds": 1.8621667921543121 } ], "all_smoke_checks_pass": true, "device": { "device_name": "Apple M5 Max", "max_recommended_working_set_size": 115448725504, "memory_size": 137438953472, "architecture": "applegpu_g17s", "max_buffer_length": 86586540032, "resource_limit": 499000 }, "mlx_version": "0.32.4", "mlx_vlm_version": "0.7.10", "converter_revision": "6d9ba328abbd6129c7461eda86280aa9fbfd9986", "python": "3.12.15", "peak_active_gb": 42.345716936, "seconds": 10.483687666011974, "swap_before": "vm.swapusage: total = 10240.00M used = 8736.50M free = 1503.50M (encrypted)", "swap_after": "vm.swapusage: total = 10240.00M used = 8736.50M free = 1503.50M (encrypted)", "limits": "Three synthetic smoke prompts and finite/cache checks; not a quality benchmark or proof of parity to BF16/Transformers." }