{ "artifacts": [ { "file": "model.litertlm", "sha256": "f7170ef16d28dd4723385dfb1afca666b4f902e93dc89dc5c1a2a36f04e05fba", "size_mb": 2232.098 } ], "benchmarks": [], "conversion": { "command": "python scripts/extract_text_backbone.py mistralai/Ministral-3-3B-Reasoning-2512 src_models/ministral-3-3b-reasoning-text && EXTERNALIZE_EMBEDDER=1 FORCE_SPM=1 CACHE=4096 python scripts/export_simple_template.py src_models/ministral-3-3b-reasoning-text out/ministral-3-3b-reasoning-ext templates/mistral_simple.jinja BOCTAV4", "quantization": "int4 blockwise-32 symmetric + OCTAV clipping; tied embedding/lm_head int8 (BOCTAV4)", "tool": "litert-torch (via litertlm-convert scripts/export_simple_template.py)", "tool_version": "0.10.0 (editable dev checkout 115a136 + local patches)" }, "cross_runtime": [], "delegation": null, "device": { "records": [ { "device": "mac-studio-m4-max", "run": { "accelerator": "gpu", "context_length": null, "date": "2026-08-12", "decode_tokens_per_s": 94.65, "delegated_ops": null, "env": { "device": "Mac Studio (M4 Max)", "machine_label": "mac-studio-m4-max", "os_build": null, "runtime": "litert-lm", "runtime_version": "0.16.0", "soc": "Apple M4 Max", "vendor_sdk": null }, "error": null, "evidence": [ "results block: prefill=1234.34 decode=94.65 tokens/s, init=2.5757 s" ], "failure_class": null, "full_delegation": null, "latency_p50_ms": null, "loads": true, "max_abs_diff": null, "max_rel_diff": null, "metrics": { "decode_tokens": 256.0, "init_s": 2.5757, "max_num_tokens": 1024.0, "prefill_tokens": 256.0 }, "output_match": null, "peak_mem_mb": null, "prefill_tokens_per_s": 1234.34, "provenance": "measured", "runs": true, "total_ops": null, "ttft_ms": 222.0 }, "source": "data/device_runs/0.16.0/2026-08-12/ministral3-3b-reasoning__mac-studio-m4-max.json" }, { "device": "raspberry-pi-5", "run": { "accelerator": "cpu", "context_length": null, "date": "2026-09-01", "decode_tokens_per_s": 1.9, "delegated_ops": null, "env": { "device": "Raspberry Pi 5 Model B Rev 1.1", "machine_label": "raspberry-pi-5-cooled-52c", "os_build": "Linux-6.18.34+rpt-rpi-2712-aarch64-with-glibc2.41", "runtime": "litert-lm", "runtime_version": "0.16.1", "soc": "Broadcom BCM2712", "vendor_sdk": null }, "error": null, "evidence": [ "pi5 LLM sweep row: `litert-lm benchmark --backend cpu --cpu-thread-count 4 -p 256 -d 256 --runs 1 --cache memory` (wave-2 driver pi5_llm_bench.py; --cache memory rather than the house --cache no, which OOM-kills every >=1.2B file on the 8 GB Pi — equivalence measured on granite-350m int8, +2-3%), 3 invocations per file with cool-down to <=52 C between them, vcgencmd measure_temp + get_throttled logged per invocation, peak RSS polled from /proc; throughput = median of the three invocations (spread in metrics); a row counts as measured only when the real-generation gate (`litert-lm run`, degenerate-output check) passed and every invocation exited 0 with get_throttled 0x0", "versions: cpu=Raspberry Pi 5 Model B Rev 1.1, litert-lm=0.16.1, litert-lm-api=0.16.1, platform=Linux-6.18.34+rpt-rpi-2712-aarch64-with-glibc2.41, python=3.13.5", "cache mode 'memory'; -p 256 -d 256 --runs 1 --cpu-thread-count 4", "gate ('What is 17 plus 26? Answer with the number only.'): status pass, exit 0, wall 191.0 s, output head '[thought] Okay, the question is asking for the sum of 17 and 26. Let me break th'", "invocation 0: exit=0 wall_s=404.3 temp 49.4->53.2C throttled=0x0 prefill_tps=14.04 decode_tps=1.9 ttft_s=21.27 init_s=56.8546 peak_rss_mb=3794", "invocation 1: exit=0 wall_s=402.3 temp 50.5->52.1C throttled=0x0 prefill_tps=14.06 decode_tps=1.91 ttft_s=20.8273 init_s=55.2938 peak_rss_mb=3794", "invocation 2: exit=0 wall_s=402.3 temp 51.0->53.8C throttled=0x0 prefill_tps=13.91 decode_tps=1.9 ttft_s=20.9882 init_s=55.1799 peak_rss_mb=3794" ], "failure_class": null, "full_delegation": null, "latency_p50_ms": null, "loads": true, "max_abs_diff": null, "max_rel_diff": null, "metrics": { "decode_tokens": 256.0, "decode_tps_max": 1.91, "decode_tps_min": 1.9, "init_s": 55.2938, "init_s_max": 56.8546, "init_s_min": 55.1799, "invocations": 3.0, "prefill_tokens": 256.0, "prefill_tps_max": 14.06, "prefill_tps_min": 13.91, "runs_per_invocation": 1.0, "threads": 4.0, "ttft_s_max": 21.27, "ttft_s_min": 20.8273 }, "output_match": null, "peak_mem_mb": 3794.0, "prefill_tokens_per_s": 14.04, "provenance": "measured", "runs": true, "total_ops": null, "ttft_ms": 20988.2 }, "source": "data/device_runs/0.16.1/2026-09-01/ministral3-3b-reasoning__raspberry-pi-5.json" } ] }, "model": { "family": "ministral", "id": "ministral3-3b-reasoning", "license": "apache-2.0", "source_url": "https://huggingface.co/mistralai/Ministral-3-3B-Reasoning-2512", "task": "text-generation" }, "pitfalls": [ "iOS: a single TFLite weight section over 2 GiB fails to mmap (engine creation error 'Failed to map section') — the tied embedding is externalized (externalize_embedder=True) so every section stays under 2 GiB and the model loads on iPhone.", "Must be exported with the native Mistral [INST] template and real EOS , not ChatML — Mistral's tokenizer has no <|im_end|> token, so a ChatML export never hits a registered stop and runs away after the answer.", "Android GPU needs roughly 2x the model size in RAM (weights plus the ML Drift GPU weight cache); GPU is only offered on ~12 GB+ devices — on an 8 GB phone only CPU is selectable.", "Reasoning model: give it a generous max-tokens budget (GSM8K was scored at max-tokens 2048; scoring at 512 falsely penalises it).", "Text-only conversion: the Pixtral vision tower is dropped before export (strict missing=0 / unexpected=0 weight check)." ], "schema_version": "1.2" }