{
  "artifacts": [
    {
      "file": "LFM2.5-350M_int4.litertlm",
      "sha256": "272b0a1ca55f8b39fe636927deffe4049feb93b5eda7c05a5dbbb3141d6d53f8",
      "size_mb": 236.271
    }
  ],
  "benchmarks": [],
  "conversion": {
    "command": "LiquidAI/LFM2.5-350M rev 9e6c6ccf: unquantized export (11 prefill signatures, 1024 down to 1, plus decode, over a 4096-position KV cache; 22 state buffers: 10 ShortConv states, 6 key and 6 value caches) -> post-hoc int4 quantization -> a scan of the int4 block scales (none of the 92 blockwise tensors has a zero or subnormal scale) -> metadata repack with LiteRT-LM's lfm2 configuration (card L110-L111)",
    "quantization": "int4 blockwise-32 + OCTAV on linears, int8 lm_head + embedding (card L28)",
    "tool": "litert-torch 0.9.4 (released wheels); repro = hf-to-litertlm lfm_work/convert_lfm25_350m.sh (card L110)",
    "tool_version": "litert-torch 0.9.4, litert-converter 0.4.0, ai-edge-quantizer 0.9.0, ai-edge-litert 2.2.0, transformers 5.14.1, torch 2.13.0 (released wheels); packed by the litert-lm 0.18.0 CLI"
  },
  "cross_runtime": [],
  "delegation": null,
  "device": {
    "records": [
      {
        "device": "galaxy-s26",
        "run": {
          "accelerator": "cpu",
          "context_length": null,
          "date": "2026-10-12",
          "decode_tokens_per_s": 90.63,
          "delegated_ops": 788,
          "env": {
            "device": "Galaxy S26 (SM-S942Q)",
            "machine_label": "galaxy-s26-aab1",
            "os_build": "Android 16",
            "runtime": "litert-lm",
            "runtime_version": "prebuilt-adac974b",
            "soc": "Qualcomm SM8850",
            "vendor_sdk": null
          },
          "error": null,
          "evidence": [
            "binary: Google's prebuilt litert_lm_advanced_main + 6 .so (gs://litert/binaries/latest/android_arm64/litert_lm/, object dated 2026-09-18, 40,382,152 B, sha256 adac974b…)",
            "r1 hold 2026-10-12 15:55:45-16:03:26 JST: [15:57:05] bundle on device: 272b0a1ca55f8b39fe636927deffe4049feb93b5eda7c05a5dbbb3141d6d53f8 (247747696 B, push 6.3 s) = Mac",
            "bench args: --backend=cpu --model_path=/data/local/tmp/aa1_350m/LFM2.5-350M_int4.litertlm --benchmark=true --benchmark_prefill_tokens=1024 --benchmark_decode_tokens=256 --max_num_tokens=2048 --num_iterations=3 --report_peak_memory_footprint=true --metric_proto_file_path=/data/local/tmp/aa1_350m/metrics.pb",
            "VERBOSE: Replacing 487 out of 560 node(s) with delegate (TfLiteXNNPackDelegate) node, yielding 88 partitions for subgraph 0 (prefill_1024).",
            "VERBOSE: Replacing 455 out of 519 node(s) with delegate (TfLiteXNNPackDelegate) node, yielding 92 partitions for subgraph 11 (decode).",
            "VERBOSE: Replacing 788 out of 818 node(s) with delegate (TfLiteXNNPackDelegate) node, yielding 14 partitions for subgraph 0 (prefill_1024).",
            "VERBOSE: Replacing 737 out of 756 node(s) with delegate (TfLiteXNNPackDelegate) node, yielding 14 partitions for subgraph 11 (decode).",
            "[15:58:35] parsed: s26_int4_cpu_timing cpu: gate PASS 1/1 | bench prefill 875.875 decode 90.63 init 602.89 cold 815.6 peak 1120.91 capped 5/7 | answers: 'The capital of France is Paris.'",
            "timing run prompt process: \"The capital of France is Paris.\"",
            "check run: 2 of 2 prompt processes answered with the expected word (\"The capital of France is Paris.\")"
          ],
          "failure_class": null,
          "full_delegation": false,
          "latency_p50_ms": null,
          "loads": true,
          "max_abs_diff": null,
          "max_rel_diff": null,
          "metrics": {
            "boot_epoch_s": 1791590919.2,
            "capped_samples": 5.0,
            "cold_init_ms": 815.6,
            "cpu_capped_samples": 5.0,
            "decode_tok_s_it1": 101.67,
            "decode_tok_s_it2": 97.12,
            "decode_tok_s_it3": 84.14,
            "decode_tokens": 256.0,
            "init_ms": 602.89,
            "iterations": 3.0,
            "kgsl_ceiling_min_mhz": 1300.0,
            "kgsl_temp_max_c": 64.3,
            "max_num_tokens": 2048.0,
            "memavail_at_bench_start_kib": 7430480.0,
            "memavail_min_kib": 6566000.0,
            "prefill_tok_s_it1": 848.38,
            "prefill_tok_s_it2": 948.08,
            "prefill_tok_s_it3": 803.67,
            "prefill_tokens": 1024.0,
            "sampler_samples": 7.0,
            "skin_c_at_bench_start": 37.9,
            "t_1024_256_s": 3.9938,
            "thermal_status_at_bench_start": 0.0,
            "ttft_s_it1": 1.22,
            "ttft_s_it2": 1.09,
            "ttft_s_it3": 1.29
          },
          "output_match": null,
          "peak_mem_mb": 1120.91,
          "prefill_tokens_per_s": 875.875,
          "provenance": "measured",
          "runs": true,
          "total_ops": 818,
          "ttft_ms": 1190.0
        },
        "source": "data/device_runs/prebuilt-adac974b/2026-10-12/lfm2.5-350m-int4__galaxy-s26.json"
      },
      {
        "device": "galaxy-s26",
        "run": {
          "accelerator": "gpu",
          "context_length": null,
          "date": "2026-10-12",
          "decode_tokens_per_s": 106.86,
          "delegated_ops": 560,
          "env": {
            "device": "Galaxy S26 (SM-S942Q)",
            "machine_label": "galaxy-s26-aab1",
            "os_build": "Android 16",
            "runtime": "litert-lm",
            "runtime_version": "prebuilt-adac974b",
            "soc": "Qualcomm SM8850",
            "vendor_sdk": null
          },
          "error": null,
          "evidence": [
            "binary: Google's prebuilt litert_lm_advanced_main + 6 .so (gs://litert/binaries/latest/android_arm64/litert_lm/, object dated 2026-09-18, 40,382,152 B, sha256 adac974b…)",
            "r1 hold 2026-10-12 15:55:45-16:03:26 JST: [16:01:08] bundle on device: 272b0a1ca55f8b39fe636927deffe4049feb93b5eda7c05a5dbbb3141d6d53f8 (247747696 B, push 0 s) = Mac",
            "bench args: --backend=gpu --model_path=/data/local/tmp/aa1_350m/LFM2.5-350M_int4.litertlm --benchmark=true --benchmark_prefill_tokens=1024 --benchmark_decode_tokens=256 --max_num_tokens=2048 --num_iterations=3 --report_peak_memory_footprint=true --metric_proto_file_path=/data/local/tmp/aa1_350m/metrics.pb",
            "VERBOSE: Replacing 560 out of 560 node(s) with delegate (LITERT_CL) node, yielding 1 partitions for subgraph 0 (prefill_1024).",
            "VERBOSE: Replacing 560 out of 560 node(s) with delegate (LITERT_CL) node, yielding 1 partitions for subgraph 1 (prefill_512).",
            "VERBOSE: Replacing 560 out of 560 node(s) with delegate (LITERT_CL) node, yielding 1 partitions for subgraph 2 (prefill_256).",
            "VERBOSE: Replacing 560 out of 560 node(s) with delegate (LITERT_CL) node, yielding 1 partitions for subgraph 3 (prefill_128).",
            "VERBOSE: Replacing 560 out of 560 node(s) with delegate (LITERT_CL) node, yielding 1 partitions for subgraph 4 (prefill_64).",
            "VERBOSE: Replacing 560 out of 560 node(s) with delegate (LITERT_CL) node, yielding 1 partitions for subgraph 5 (prefill_32).",
            "VERBOSE: Replacing 560 out of 560 node(s) with delegate (LITERT_CL) node, yielding 1 partitions for subgraph 6 (prefill_16).",
            "VERBOSE: Replacing 560 out of 560 node(s) with delegate (LITERT_CL) node, yielding 1 partitions for subgraph 7 (prefill_8).",
            "VERBOSE: Replacing 560 out of 560 node(s) with delegate (LITERT_CL) node, yielding 1 partitions for subgraph 8 (prefill_4).",
            "VERBOSE: Replacing 560 out of 560 node(s) with delegate (LITERT_CL) node, yielding 1 partitions for subgraph 9 (prefill_2).",
            "VERBOSE: Replacing 501 out of 501 node(s) with delegate (LITERT_CL) node, yielding 1 partitions for subgraph 10 (prefill_1).",
            "VERBOSE: Replacing 519 out of 519 node(s) with delegate (LITERT_CL) node, yielding 1 partitions for subgraph 11 (decode).",
            "[16:02:39] parsed: s26_int4_gpu_timing gpu: gate PASS 1/1 | bench prefill 4391.315 decode 106.86000000000001 init 1194.9 cold 4037.92 peak 472.211 capped 0/5 | answers: 'The capital of France is Paris.'",
            "timing run prompt process: \"The capital of France is Paris.\"",
            "check run: 2 of 2 prompt processes answered with the expected word (\"The capital of France is Paris.\")"
          ],
          "failure_class": null,
          "full_delegation": true,
          "latency_p50_ms": null,
          "loads": true,
          "max_abs_diff": null,
          "max_rel_diff": null,
          "metrics": {
            "boot_epoch_s": 1791590918.9,
            "capped_samples": 0.0,
            "cold_init_ms": 4037.92,
            "cpu_capped_samples": 4.0,
            "decode_tok_s_it1": 97.55,
            "decode_tok_s_it2": 107.79,
            "decode_tok_s_it3": 105.93,
            "decode_tokens": 256.0,
            "init_ms": 1194.9,
            "iterations": 3.0,
            "kgsl_ceiling_min_mhz": 1300.0,
            "kgsl_temp_max_c": 79.5,
            "max_num_tokens": 2048.0,
            "memavail_at_bench_start_kib": 7450792.0,
            "memavail_min_kib": 6293760.0,
            "prefill_tok_s_it1": 4558.61,
            "prefill_tok_s_it2": 4361.69,
            "prefill_tok_s_it3": 4420.94,
            "prefill_tokens": 1024.0,
            "sampler_samples": 5.0,
            "skin_c_at_bench_start": 37.9,
            "t_1024_256_s": 2.6288,
            "thermal_status_at_bench_start": 0.0,
            "ttft_s_it1": 0.23,
            "ttft_s_it2": 0.24,
            "ttft_s_it3": 0.24
          },
          "output_match": null,
          "peak_mem_mb": 472.211,
          "prefill_tokens_per_s": 4391.315,
          "provenance": "measured",
          "runs": true,
          "total_ops": 560,
          "ttft_ms": 240.0
        },
        "source": "data/device_runs/prebuilt-adac974b/2026-10-12/lfm2.5-350m-int4__galaxy-s26.json"
      },
      {
        "device": "mac-studio-m4-max",
        "run": {
          "accelerator": "cpu",
          "context_length": null,
          "date": "2026-10-12",
          "decode_tokens_per_s": 123.11,
          "delegated_ops": null,
          "env": {
            "device": "Mac Studio (M4 Max)",
            "machine_label": "mac-studio-m4-max",
            "os_build": "macOS 27.0",
            "runtime": "litert-lm",
            "runtime_version": "0.18.0",
            "soc": "Apple M4 Max",
            "vendor_sdk": null
          },
          "error": null,
          "evidence": [
            "command (tools/bench_mac.sh): litert-lm benchmark LFM2.5-350M_int4.litertlm --backend cpu -p 1024 -d 256 --max-num-tokens 2048 --runs 3 --cache no",
            "Benchmarking model: LFM2.5-350M_int4.litertlm | Backend                    : cpu | Cache                      : no | Prefill speed:        1195.41 tokens/s | Decode speed:         123.11 tokens/s | Init time:            1.1431 s | Time to first token:  0.8649 s",
            "2026-10-12 15:31:34 JST cpu int4 exit 0"
          ],
          "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": 1.1431,
            "iterations": 3.0,
            "max_num_tokens": 2048.0,
            "passes": 1.0,
            "prefill_tokens": 1024.0,
            "t_1024_256_s": 2.9361
          },
          "output_match": null,
          "peak_mem_mb": null,
          "prefill_tokens_per_s": 1195.41,
          "provenance": "measured",
          "runs": true,
          "total_ops": null,
          "ttft_ms": 864.9
        },
        "source": "data/device_runs/0.18.0/2026-10-12/lfm2.5-350m-int4__mac-studio-m4-max.json"
      },
      {
        "device": "mac-studio-m4-max",
        "run": {
          "accelerator": "gpu",
          "context_length": null,
          "date": "2026-10-12",
          "decode_tokens_per_s": 425.85,
          "delegated_ops": null,
          "env": {
            "device": "Mac Studio (M4 Max)",
            "machine_label": "mac-studio-m4-max",
            "os_build": "macOS 27.0",
            "runtime": "litert-lm",
            "runtime_version": "0.18.0",
            "soc": "Apple M4 Max",
            "vendor_sdk": null
          },
          "error": null,
          "evidence": [
            "command (tools/bench_mac.sh): litert-lm benchmark LFM2.5-350M_int4.litertlm --backend gpu -p 1024 -d 256 --max-num-tokens 2048 --runs 3 --cache no",
            "Benchmarking model: LFM2.5-350M_int4.litertlm | Backend                    : gpu | Cache                      : no | Prefill speed:        13965.02 tokens/s | Decode speed:         425.85 tokens/s | Init time:            1.1679 s | Time to first token:  0.0757 s",
            "2026-10-12 15:37:03 JST gpu int4 exit 0",
            "2026-10-12 15:37:14 JST gpu int4 verbose exit 0 validation_error=0 shape_mismatch=0 replacing=24"
          ],
          "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": 1.1679,
            "iterations": 3.0,
            "max_num_tokens": 2048.0,
            "passes": 1.0,
            "prefill_tokens": 1024.0,
            "t_1024_256_s": 0.6745
          },
          "output_match": null,
          "peak_mem_mb": null,
          "prefill_tokens_per_s": 13965.02,
          "provenance": "measured",
          "runs": true,
          "total_ops": null,
          "ttft_ms": 75.7
        },
        "source": "data/device_runs/0.18.0/2026-10-12/lfm2.5-350m-int4__mac-studio-m4-max.json"
      },
      {
        "device": "pixel-8a",
        "run": {
          "accelerator": "cpu",
          "context_length": null,
          "date": "2026-10-12",
          "decode_tokens_per_s": 51.405,
          "delegated_ops": 788,
          "env": {
            "device": "Pixel 8a",
            "machine_label": "pixel-8a-aab1",
            "os_build": "Android 16",
            "runtime": "litert-lm",
            "runtime_version": "prebuilt-adac974b",
            "soc": "Tensor G3",
            "vendor_sdk": null
          },
          "error": null,
          "evidence": [
            "binary: Google's prebuilt litert_lm_advanced_main + 6 .so (gs://litert/binaries/latest/android_arm64/litert_lm/, object dated 2026-09-18, 40,382,152 B, sha256 adac974b…)",
            "r1 hold 2026-10-12 15:55:45-16:00:35 JST: [15:57:48] bundle on device: 272b0a1ca55f8b39fe636927deffe4049feb93b5eda7c05a5dbbb3141d6d53f8 (247747696 B, push 0 s) = Mac",
            "bench args: --backend=cpu --model_path=/data/local/tmp/aa1_350m/LFM2.5-350M_int4.litertlm --benchmark=true --benchmark_prefill_tokens=1024 --benchmark_decode_tokens=256 --max_num_tokens=2048 --num_iterations=3 --report_peak_memory_footprint=true --metric_proto_file_path=/data/local/tmp/aa1_350m/metrics.pb",
            "VERBOSE: Replacing 487 out of 560 node(s) with delegate (TfLiteXNNPackDelegate) node, yielding 88 partitions for subgraph 0 (prefill_1024).",
            "VERBOSE: Replacing 455 out of 519 node(s) with delegate (TfLiteXNNPackDelegate) node, yielding 92 partitions for subgraph 11 (decode).",
            "VERBOSE: Replacing 788 out of 818 node(s) with delegate (TfLiteXNNPackDelegate) node, yielding 14 partitions for subgraph 0 (prefill_1024).",
            "VERBOSE: Replacing 737 out of 756 node(s) with delegate (TfLiteXNNPackDelegate) node, yielding 14 partitions for subgraph 11 (decode).",
            "[15:59:02] parsed: p8a_int4_cpu_timing cpu: gate PASS 1/1 | bench prefill 342.115 decode 51.405 init 1245.32 cold 1457.95 peak 835.961 capped 0/10 | answers: 'The capital of France is Paris.'",
            "timing run prompt process: \"The capital of France is Paris.\"",
            "check run: 1 of 1 prompt processes answered with the expected word (\"The capital of France is Paris.\")"
          ],
          "failure_class": null,
          "full_delegation": false,
          "latency_p50_ms": null,
          "loads": true,
          "max_abs_diff": null,
          "max_rel_diff": null,
          "metrics": {
            "boot_epoch_s": 1791442100.14,
            "capped_samples": 0.0,
            "cold_init_ms": 1457.95,
            "cpu_capped_samples": 0.0,
            "decode_tok_s_it1": 50.35,
            "decode_tok_s_it2": 52.27,
            "decode_tok_s_it3": 50.54,
            "decode_tokens": 256.0,
            "init_ms": 1245.32,
            "iterations": 3.0,
            "max_num_tokens": 2048.0,
            "memavail_at_bench_start_kib": 3842512.0,
            "memavail_min_kib": 3283672.0,
            "prefill_tok_s_it1": 331.5,
            "prefill_tok_s_it2": 348.44,
            "prefill_tok_s_it3": 335.79,
            "prefill_tokens": 1024.0,
            "sampler_samples": 10.0,
            "t_1024_256_s": 7.9732,
            "thermal_status_at_bench_start": 0.0,
            "ttft_s_it1": 3.11,
            "ttft_s_it2": 2.96,
            "ttft_s_it3": 3.07
          },
          "output_match": null,
          "peak_mem_mb": 835.961,
          "prefill_tokens_per_s": 342.115,
          "provenance": "measured",
          "runs": true,
          "total_ops": 818,
          "ttft_ms": 3015.0
        },
        "source": "data/device_runs/prebuilt-adac974b/2026-10-12/lfm2.5-350m-int4__pixel-8a.json"
      },
      {
        "device": "pixel-8a",
        "run": {
          "accelerator": "gpu",
          "context_length": null,
          "date": "2026-10-12",
          "decode_tokens_per_s": 50.055,
          "delegated_ops": 560,
          "env": {
            "device": "Pixel 8a",
            "machine_label": "pixel-8a-aab1",
            "os_build": "Android 16",
            "runtime": "litert-lm",
            "runtime_version": "prebuilt-adac974b",
            "soc": "Tensor G3",
            "vendor_sdk": null
          },
          "error": null,
          "evidence": [
            "binary: Google's prebuilt litert_lm_advanced_main + 6 .so (gs://litert/binaries/latest/android_arm64/litert_lm/, object dated 2026-09-18, 40,382,152 B, sha256 adac974b…)",
            "r2 hold 2026-10-12 16:05:41-16:06:56 JST: [16:05:50] bundle on device: 272b0a1ca55f8b39fe636927deffe4049feb93b5eda7c05a5dbbb3141d6d53f8 (247747696 B, push 5.3 s) = Mac",
            "bench args: --backend=gpu --model_path=/data/local/tmp/aa1_350m/LFM2.5-350M_int4.litertlm --benchmark=true --benchmark_prefill_tokens=1024 --benchmark_decode_tokens=256 --max_num_tokens=2048 --num_iterations=3 --report_peak_memory_footprint=true --metric_proto_file_path=/data/local/tmp/aa1_350m/metrics.pb",
            "VERBOSE: Replacing 560 out of 560 node(s) with delegate (LITERT_CL) node, yielding 1 partitions for subgraph 0 (prefill_1024).",
            "VERBOSE: Replacing 560 out of 560 node(s) with delegate (LITERT_CL) node, yielding 1 partitions for subgraph 1 (prefill_512).",
            "VERBOSE: Replacing 560 out of 560 node(s) with delegate (LITERT_CL) node, yielding 1 partitions for subgraph 2 (prefill_256).",
            "VERBOSE: Replacing 560 out of 560 node(s) with delegate (LITERT_CL) node, yielding 1 partitions for subgraph 3 (prefill_128).",
            "VERBOSE: Replacing 560 out of 560 node(s) with delegate (LITERT_CL) node, yielding 1 partitions for subgraph 4 (prefill_64).",
            "VERBOSE: Replacing 560 out of 560 node(s) with delegate (LITERT_CL) node, yielding 1 partitions for subgraph 5 (prefill_32).",
            "VERBOSE: Replacing 560 out of 560 node(s) with delegate (LITERT_CL) node, yielding 1 partitions for subgraph 6 (prefill_16).",
            "VERBOSE: Replacing 560 out of 560 node(s) with delegate (LITERT_CL) node, yielding 1 partitions for subgraph 7 (prefill_8).",
            "VERBOSE: Replacing 560 out of 560 node(s) with delegate (LITERT_CL) node, yielding 1 partitions for subgraph 8 (prefill_4).",
            "VERBOSE: Replacing 560 out of 560 node(s) with delegate (LITERT_CL) node, yielding 1 partitions for subgraph 9 (prefill_2).",
            "VERBOSE: Replacing 501 out of 501 node(s) with delegate (LITERT_CL) node, yielding 1 partitions for subgraph 10 (prefill_1).",
            "VERBOSE: Replacing 519 out of 519 node(s) with delegate (LITERT_CL) node, yielding 1 partitions for subgraph 11 (decode).",
            "[16:06:43] parsed: p8a_int4_gpu_timing gpu: gate PASS 1/1 | bench prefill 1245.8649999999998 decode 50.055 init 1833.5 cold 6078.84 peak 872.18 capped 0/8 | answers: 'The capital of France is Paris.'",
            "timing run prompt process: \"The capital of France is Paris.\"",
            "check run: 2 of 2 prompt processes answered with the expected word (\"The capital of France is Paris.\")"
          ],
          "failure_class": null,
          "full_delegation": true,
          "latency_p50_ms": null,
          "loads": true,
          "max_abs_diff": null,
          "max_rel_diff": null,
          "metrics": {
            "boot_epoch_s": 1791442099.96,
            "capped_samples": 0.0,
            "cold_init_ms": 6078.84,
            "cpu_capped_samples": 0.0,
            "decode_tok_s_it1": 43.51,
            "decode_tok_s_it2": 51.56,
            "decode_tok_s_it3": 48.55,
            "decode_tokens": 256.0,
            "init_ms": 1833.5,
            "iterations": 3.0,
            "max_num_tokens": 2048.0,
            "memavail_at_bench_start_kib": 3819608.0,
            "memavail_min_kib": 2809764.0,
            "prefill_tok_s_it1": 1198.11,
            "prefill_tok_s_it2": 1244.11,
            "prefill_tok_s_it3": 1247.62,
            "prefill_tokens": 1024.0,
            "sampler_samples": 8.0,
            "t_1024_256_s": 5.9363,
            "thermal_status_at_bench_start": 0.0,
            "ttft_s_it1": 0.88,
            "ttft_s_it2": 0.84,
            "ttft_s_it3": 0.84
          },
          "output_match": null,
          "peak_mem_mb": 872.18,
          "prefill_tokens_per_s": 1245.865,
          "provenance": "measured",
          "runs": true,
          "total_ops": 560,
          "ttft_ms": 840.0
        },
        "source": "data/device_runs/prebuilt-adac974b/2026-10-12/lfm2.5-350m-int4__pixel-8a.json"
      }
    ]
  },
  "model": {
    "family": "lfm2.5",
    "id": "lfm2.5-350m-int4",
    "license": "other (lfm1.0)",
    "source_url": "https://huggingface.co/litert-community/LFM2.5-350M",
    "task": "text-generation"
  },
  "pitfalls": [
    "Runtime: tested on litert-lm 0.18.0 on a Mac; this file was not tried on litert-lm 0.17.1, where the int8 file of the same conversion answered as if no question had been asked (the runtime prefilled 9 tokens for a one-question turn where 0.18.0 prefilled 16) (card L21).",
    "Template: Liquid AI's chat template fails at template creation on litert-lm 0.18.0 ({% generation %} markers), so the bundle carries LiteRT-LM's lfm2 template (byte-identical to models/lfm2/chat_template.jinja at commit 86a7e6d0), which reads message content as a list of parts; stops <|im_end|> and <|endoftext|>, start token <|startoftext|>; marked non-thinking, with the configuration's thought channel kept (card L71, L112).",
    "One Engine per conversation: on the int8 file, an Engine that had already served one conversation changed 2 of the next conversation's 3 answers (Python API, CPU, greedy, 3 of 3 processes); a related open issue is LiteRT-LM #3165; the card advises a new Engine for each conversation (card L75).",
    "Conversation preface with tools and no messages: LiteRT-LM renders the chat template with an empty message list, and the template reads messages[0] without checking that the list is non-empty, so the render fails with 'Failed to apply template: undefined value' when the preface is prefilled at creation (prefillPrefaceOnInit, default false; creating the conversation fails) and in the preface render (renderPrefaceIntoString and its C++ and C API forms); the same with extra context (such as enable_thinking) and no messages, and in the preface render with an empty preface (run in the runtime on one Qwen3 template; the other templates rendered with minijinja 2.14 and jinja2); read in the source of LiteRT-LM 0.17.0, 0.18.1 and main, measured with the 0.18.1 PyPI package through its C API on these files' template text; the default path, where the first message is rendered with the preface, is not affected (the 0.18.1 Python API and CLI expose neither the flag nor the preface render); a system message or any initial message in the conversation config avoids it (card L116).",
    "Sampler: the bundle's default sampler is temperature 0.1, top-k 50, top-p 1.0 (Liquid AI's recommended repetition_penalty 1.05 is not carried); `run` uses it unless --top-k, --top-p, --temperature or --seed is passed; --top-k 1 gives greedy output (card L69).",
    "Cache: without --cache, `run` writes a compiled-weights cache next to the model (231 MB for this file on CPU); without --backend, `run` uses the CPU (card L67).",
    "Tool calling is declared (model type lfm2); after a tool call the bundle template and Liquid AI's differ in two places (call arguments in double quotes, the tool result wrapped in [ … ]); on litert-lm 0.18.0 (Python API, CPU, greedy) the file returned a parsed get_weather call with {\"city\": \"Paris\"} (card L73).",
    "Which file is faster differs by device: int4 on the Galaxy S26 GPU (2.63 s vs int8 2.94 s for a 1024-token prefill plus 256 decoded tokens), int8 on the Pixel 8a GPU (5.77 vs 5.94 s: there int4 decodes at 50.06 vs 47.9 tok/s and prefills at 1245.9 vs 2410.6 tok/s), within 1 % on the Mac GPU (0.67 vs 0.68 s); 38.6 % smaller than the int8 file; not measured on iOS, not run on the NPU (card L30-L39, L101-L102, L106).",
    "Galaxy S26 CPU: the row ran with the clock lowering during it (lowered in 5 of 7 samples, read every 2 s); decode fell across the three iterations (101.67 -> 97.12 -> 84.14 tok/s) (card L97).",
    "Accuracy: IFEval-lite (first 120 rows of google/IFEval whose instruction types are checked mechanically, prompt-level strict, greedy; comparable only within that table) bf16 81 (67.5 %), this file 79 (65.8 %) on CPU and 78 (65.0 %) on the Mac GPU; 8-question check 7/8 on CPU and on GPU (missing \"Which is larger: 0.9 or 0.11?\" -> \"0.11 is larger than 0.9.\"), bf16 7/8; in a three-turn CPU conversation it answered with a different word (\"large\") than bf16 and completed all three turns; GSM8K 8 of 100 (bf16 2) (card L43, L45, L49-L57).",
    "License: the LFM Open License v1.0's section 5 conditions commercial use on an annual revenue below 10 million US dollars (card L120)."
  ],
  "schema_version": "1.2"
}
