{
  "artifacts": [
    {
      "file": "WeVisOne-VL-4B_int4.litertlm",
      "sha256": "6ae2e63e83516bf8a8f382c9938c0be8f716cba2a66b8bbcc22991e9123c1e48",
      "size_mb": 3345.455
    }
  ],
  "benchmarks": [],
  "conversion": {
    "command": "tencent/WeVisOne-VL-4B float32 release rev 4a7bed92 (its bfloat16 release 6b44a458 holds the same weights) -> the WeVisOne-VL-2B recipe of hf-to-litertlm at the 4B's dimensions, with the int4 decoder and the GPU graph step (L160); a single-picture fast_vlm file with 2,048 tokens of context for the prompt and the answer (L33); re-packed with LiteRT external weights on 2026-11-12, the same weight bytes (L164)",
    "quantization": "decoder int4 blockwise 32 with one scale per block, every block scale at or above 2^-14, its output layer included; token embedder int8 (one scale per output channel); vision encoder and adapter float16; float32 activations declared for the decoder (L30, L156)",
    "tool": "litert-torch 0.10.0 (L22); the decoder carries Qwen3-VL's three DeepStack features in each image token's row and computes the three-axis positions in the graph, the vision tower is rewritten in plain PyTorch for a fixed 512 x 512 input without a gather op, and after the export hf-to-litertlm's gpu_graph_retrofit.py rewrote the decoder's KV-cache writes and attention products into LiteRT-LM's GPU composites (odml.cache_update, odml.runtime_bmm); one fast_vlm bundle with the source repository's own tokenizer.json byte for byte (L33, L153-L158); since 2026-11-12 the decoder's and the token embedder's weights sit in TFLiteWeights sections at the end of the file (LiteRT's external weights) and the vision encoder and adapter keep theirs in the graph (L33, L164); recipe = hf-to-litertlm REPRODUCE.md's WeVisOne-VL-4B entry, which rebuilds the earlier int4 file (revision ee03d010, before its weights moved) with every section equal to that file's (L160)",
    "tool_version": "litert-torch 0.10.0 (L22), ai-edge-quantizer 0.10.0 (L156); the phone numbers on LiteRT-LM 0.19.1's litert_lm_advanced_main built from the tag for android_arm64 (L44, L77); the Mac table on the PyPI litert-lm 0.19.1 package (L103)"
  },
  "cross_runtime": [],
  "delegation": null,
  "device": {
    "records": [
      {
        "device": "galaxy-s26",
        "run": {
          "accelerator": "gpu",
          "context_length": 2048,
          "date": "2026-11-10",
          "decode_tokens_per_s": null,
          "delegated_ops": null,
          "env": {
            "device": "Galaxy S26 (SM-S942Q)",
            "machine_label": "galaxy-s26-wevisone-devicebench",
            "os_build": "Android 16 (BP4A.251205.006.S942QOPS1AZH9)",
            "runtime": "litert-lm",
            "runtime_version": "0.19.1",
            "soc": "Qualcomm SM8850 (Adreno GPU)",
            "vendor_sdk": null
          },
          "error": null,
          "evidence": [
            "binary: LiteRT-LM v0.19.1 (tag v0.19.1 = 7a03724cece4) android_arm64: litert_lm_advanced_main sha256 5d869be730622b85 (tag build of 2026-11-07, GPU accelerator statically linked) and, on bin= runs, litert_lm_advanced_main_pregpu sha256 f3f2b76a6ba3cf38 (the same commit loading the tag prebuilt libLiteRtGpuAccelerator.so, 2026-11-08) + the tag 5 prebuilt .so (no TopK sampler .so in the v0.19.1 prebuilt); \"litert_lm_advanced_main = the recorded 5d869be730622b854fb87bf74205c7b083edece8197c9279317749dfd26772b7\" (the kit checked against its SHA256SUMS at the chain's start)",
            "bundle: WeVisOne-VL-4B_int4.litertlm = the wv4b_i4b32_act32_rg.litertlm (R-G, built 2026-10-27 20:14), 3507874304 B, sha256 1c14c17a7e8bd688d888300a68e2f7512b255a275a444a3f22bde395cad4b12f (the Hub tree's LFS oid at the read below); pushed to the phone and its sha256 read back against the models.tsv: \"I4B32act32rg: 3507874304 B (sha256 read back on the phone after the push) wv4b_i4b32_act32_rg.litertlm\"; every run below records model_sha256 = artifact_sha256",
            "protocol: a benchmark script, one litert_lm_advanced_main process per run, the phone-side memory guard devside_watch.sh, a ready check (thermal, CPU cap, kgsl level, rest) before each run; paths below are relative",
            "output_match = the card's pass rule (the work line, fixed 2026-10-30 before the 4B files were scored): in each of four groups (boxes, counts, points, the no-target question) at most one fewer right than the same file on LiteRT-LM 0.18.1 on the same phone, OCR CER at most 0.02 higher, at least as many text questions right, 'Paris' to the capital of France; verdict PASS {\"box\": true, \"count\": true, \"point\": true, \"none\": true, \"ocr\": true, \"text\": true, \"france\": true}",
            "known_* = the questions with a known answer (form_score; the reference run and the float32 original score the same)",
            "reference = the source package Tencent/WeVisOne-VL @ 40f735b8 in float32 on the CPU, greedy, the same 512 x 512 pictures (arm B); box_matched … text_same_as_reference = the agreement with it (suites.s26_I4B32act32rg_gpu_vgpu, the r3 scorer unchanged)",
            "rows_token_identical_to_0181 / text_answers_identical_to_0181 = this run against the LiteRT-LM 0.18.1 run of 2026-10-28 on the same phone, token for token (token_vs_reference)",
            "suites.s26_I4B32act32rg_gpu_vgpu: {\"counts\": {\"box_matched\": 57, \"box_ref\": 58, \"count_equal\": 6, \"count_rows\": 6, \"point_matched\": 13, \"point_ref\": 14, \"exact_rows\": 3}, \"rates\": {\"box\": 0.9827586206896551, \"count\": 1.0, \"point\": 0.9285714285714286, \"cer\": 0.01639344262295082}, \"verdict\": {\"original\": {\"per_metric\": {\"box\": true, \"count\": true, \"point\": true, \"cer\": true}, \"pass_image\": true}, \"bf16\": {\"per_metric\": {\"box\": true, \"count\": true, \"point\": true, \"cer\": true}, \"pass_image\": true}}, \"text\": {\"rows\": 10, \"pass_fail_same\": 10, \"same_text_as_T0\": 6, \"pass\": 10}}"
          ],
          "failure_class": null,
          "full_delegation": null,
          "latency_p50_ms": null,
          "loads": true,
          "max_abs_diff": null,
          "max_rel_diff": null,
          "metrics": {
            "box_matched": 57.0,
            "box_ref": 58.0,
            "count_equal": 6.0,
            "count_rows": 6.0,
            "known_box_right": 11.0,
            "known_box_rows": 11.0,
            "known_count_right": 3.0,
            "known_count_rows": 4.0,
            "known_none_right": 1.0,
            "known_none_rows": 1.0,
            "known_ocr_cer": 0.0,
            "known_point_right": 2.0,
            "known_point_rows": 2.0,
            "known_text_right": 10.0,
            "known_text_rows": 10.0,
            "ocr_cer": 0.01639344262295082,
            "point_matched": 13.0,
            "point_ref": 14.0,
            "rows_token_identical": 3.0,
            "rows_token_identical_to_0181": 29.0,
            "text_answers_identical_to_0181": 10.0,
            "text_pass_fail_same": 10.0,
            "text_rows": 10.0,
            "text_same_as_reference": 6.0
          },
          "output_match": true,
          "peak_mem_mb": null,
          "prefill_tokens_per_s": null,
          "provenance": "measured",
          "runs": true,
          "signature": "the suite: 29 picture questions (17 pictures, 512 x 512 PNG) + 10 text questions, one process per question, greedy",
          "total_ops": null,
          "ttft_ms": null
        },
        "source": "data/device_runs/0.19.1/2026-11-10/wevisone-vl-4b-int4__galaxy-s26.json"
      },
      {
        "device": "galaxy-s26",
        "run": {
          "accelerator": "gpu",
          "context_length": 2048,
          "date": "2026-11-10",
          "decode_tokens_per_s": 15.2,
          "delegated_ops": null,
          "env": {
            "device": "Galaxy S26 (SM-S942Q)",
            "machine_label": "galaxy-s26-wevisone-devicebench",
            "os_build": "Android 16 (BP4A.251205.006.S942QOPS1AZH9)",
            "runtime": "litert-lm",
            "runtime_version": "0.19.1",
            "soc": "Qualcomm SM8850 (Adreno GPU)",
            "vendor_sdk": null
          },
          "error": null,
          "evidence": [
            "binary: LiteRT-LM v0.19.1 (tag v0.19.1 = 7a03724cece4) android_arm64: litert_lm_advanced_main sha256 5d869be730622b85 (tag build of 2026-11-07, GPU accelerator statically linked) and, on bin= runs, litert_lm_advanced_main_pregpu sha256 f3f2b76a6ba3cf38 (the same commit loading the tag prebuilt libLiteRtGpuAccelerator.so, 2026-11-08) + the tag 5 prebuilt .so (no TopK sampler .so in the v0.19.1 prebuilt); \"litert_lm_advanced_main = the recorded 5d869be730622b854fb87bf74205c7b083edece8197c9279317749dfd26772b7\" (the kit checked against its SHA256SUMS at the chain's start)",
            "bundle: WeVisOne-VL-4B_int4.litertlm = the wv4b_i4b32_act32_rg.litertlm (R-G, built 2026-10-27 20:14), 3507874304 B, sha256 1c14c17a7e8bd688d888300a68e2f7512b255a275a444a3f22bde395cad4b12f (the Hub tree's LFS oid at the read below); pushed to the phone and its sha256 read back against the models.tsv: \"I4B32act32rg: 3507874304 B (sha256 read back on the phone after the push) wv4b_i4b32_act32_rg.litertlm\"; every run below records model_sha256 = artifact_sha256",
            "protocol: a benchmark script, one litert_lm_advanced_main process per run, the phone-side memory guard devside_watch.sh, a ready check (thermal, CPU cap, kgsl level, rest) before each run; paths below are relative",
            "command: ./litert_lm_advanced_main --backend=gpu --model_path=/data/local/tmp/wv4b_r5/I4B32act32rg.litertlm --input_prompt_file=/data/local/tmp/wv4b_r5/prompts/vl_syn_a.txt --max_num_tokens=2048 --max_output_tokens=1024 --async=false --benchmark --benchmark_prefill_tokens=0 --benchmark_decode_tokens=0 --use_session=false --report_peak_memory_footprint=true --vision_backend=gpu",
            "prompt vl_syn_a = [image:syn_a 512 x 512 PNG] + the detection question (297 prefill tokens, 60 decoded; the run's prompts, sha256 checked on the phone at the chain's start)",
            "ttft_ms = the picture's first token = the vision encoder's time + the runtime's time to first token; medians of the two warm processes after the chain's cold one (vc, the first process of this file in the hold) built the GPU caches (key s26|I4B32act32rg|gpu|gpu|vl_timing warm_median); peak_mem_mb = the median of VmHWM / 1024",
            "{\"tag\": \"I4B32act32rg-gpu-vc\", \"status\": \"DONE\", \"cold\": true, \"cache_plan\": \"cold\", \"precision\": \"fp32\", \"ready_note\": \"clean; rested 123s since 15:53:08\", \"memory\": \"VmHWM_max_kB 3453156 VmSwap_max_kB 413532 MemAvailable_min_kB 1081888 kgsl_page_alloc_max_B 6729691136\", \"bench\": {\"ttft_s\": 3.28, \"prefill\": [[297, 92.33, 3.2168]], \"decode\": [[60, 15.01, 3.9986]], \"peak_mem_mb\": 3372.22, \"marks_s\": {\"vision_executor\": 1.305199999}}, \"init_total_ms\": 14064.61}",
            "{\"tag\": \"I4B32act32rg-gpu-vA1\", \"status\": \"DONE\", \"cold\": false, \"cache_plan\": \"keep\", \"precision\": \"fp32\", \"ready_note\": \"uncapped (thermal 1, no CPU cap, kgsl_pwrlevel 0); rested 79s since 16:04:36\", \"memory\": \"VmHWM_max_kB 3692504 VmSwap_max_kB 295480 MemAvailable_min_kB 1519444 kgsl_page_alloc_max_B 6730539008\", \"bench\": {\"ttft_s\": 3.45, \"prefill\": [[297, 87.74, 3.385]], \"decode\": [[60, 15.23, 3.9407]], \"peak_mem_mb\": 3605.96, \"marks_s\": {\"vision_executor\": 1.391035155}}, \"init_total_ms\": 11282.86}",
            "{\"tag\": \"I4B32act32rg-gpu-vA2\", \"status\": \"DONE\", \"cold\": false, \"cache_plan\": \"keep\", \"precision\": \"fp32\", \"ready_note\": \"uncapped (thermal 1, no CPU cap, kgsl_pwrlevel 0); rested 79s since 16:09:46\", \"memory\": \"VmHWM_max_kB 3698096 VmSwap_max_kB 276352 MemAvailable_min_kB 1383616 kgsl_page_alloc_max_B 6741667840\", \"bench\": {\"ttft_s\": 3.49, \"prefill\": [[297, 86.81, 3.4211]], \"decode\": [[60, 15.17, 3.9541]], \"peak_mem_mb\": 3611.42, \"marks_s\": {\"vision_executor\": 1.397318333}}, \"init_total_ms\": 11786.7}",
            "delegation: Replacing 1812 out of 1812 node(s) with delegate (LITERT_CL) node, yielding 1 partitions for subgraph 0 (prefill_1024). | Replacing 1812 out of 1812 node(s) with delegate (LITERT_CL) node, yielding 1 partitions for subgraph 1 (prefill_256). | Replacing 1812 out of 1812 node(s) with delegate (LITERT_CL) node, yielding 1 partitions for subgraph 2 (prefill_64). | Replacing 1812 out of 1812 node(s) with delegate (LITERT_CL) node, yielding 1 partitions for subgraph 3 (prefill_16). | Replacing 1812 out of 1812 node(s) with delegate (LITERT_CL) node, yielding 1 partitions for subgraph 4 (prefill_4). | Replacing 1705 out of 1705 node(s) with delegate (LITERT_CL) node, yielding 1 partitions for subgraph 5 (prefill_1). | Replacing 1737 out of 1737 node(s) with delegate (LITERT_CL) node, yielding 1 partitions for subgraph 6 (decode). | Replacing 2 out of 5 node(s) with delegate (TfLiteXNNPackDelegate) node, yielding 2 partitions for subgraph 0 (main). | Replacing 1565 out of 1565 node(s) with delegate (LITERT_CL) node, yielding 1 partitions for subgraph 0 (main). | Replacing 87 out of 111 node(s) with delegate (TfLiteXNNPackDelegate) node, yielding 3 partitions for subgraph 0 (main)."
          ],
          "failure_class": null,
          "full_delegation": null,
          "latency_p50_ms": null,
          "loads": true,
          "max_abs_diff": null,
          "max_rel_diff": null,
          "metrics": {
            "cold_decode_tok_s": 15.01,
            "cold_first_token_ms": 4585.2,
            "cold_init_total_ms": 14064.61,
            "decode_tokens": 60.0,
            "init_total_ms": 11534.78,
            "max_num_tokens": 2048.0,
            "prefill_tokens": 297.0,
            "processes": 2.0,
            "runtime_ttft_ms": 3470.0,
            "vision_encode_ms": 1394.2,
            "vmhwm_max_kb": 3698096.0
          },
          "output_match": null,
          "peak_mem_mb": 3608.69,
          "prefill_tokens_per_s": 87.275,
          "provenance": "measured",
          "runs": true,
          "signature": null,
          "total_ops": null,
          "ttft_ms": 4864.2
        },
        "source": "data/device_runs/0.19.1/2026-11-10/wevisone-vl-4b-int4__galaxy-s26.json"
      },
      {
        "device": "galaxy-s26",
        "run": {
          "accelerator": "gpu",
          "context_length": 2048,
          "date": "2026-11-10",
          "decode_tokens_per_s": 15.23,
          "delegated_ops": null,
          "env": {
            "device": "Galaxy S26 (SM-S942Q)",
            "machine_label": "galaxy-s26-wevisone-devicebench",
            "os_build": "Android 16 (BP4A.251205.006.S942QOPS1AZH9)",
            "runtime": "litert-lm",
            "runtime_version": "0.19.1",
            "soc": "Qualcomm SM8850 (Adreno GPU)",
            "vendor_sdk": null
          },
          "error": null,
          "evidence": [
            "binary: LiteRT-LM v0.19.1 (tag v0.19.1 = 7a03724cece4) android_arm64: litert_lm_advanced_main sha256 5d869be730622b85 (tag build of 2026-11-07, GPU accelerator statically linked) and, on bin= runs, litert_lm_advanced_main_pregpu sha256 f3f2b76a6ba3cf38 (the same commit loading the tag prebuilt libLiteRtGpuAccelerator.so, 2026-11-08) + the tag 5 prebuilt .so (no TopK sampler .so in the v0.19.1 prebuilt); \"litert_lm_advanced_main = the recorded 5d869be730622b854fb87bf74205c7b083edece8197c9279317749dfd26772b7\" (the kit checked against its SHA256SUMS at the chain's start)",
            "bundle: WeVisOne-VL-4B_int4.litertlm = the wv4b_i4b32_act32_rg.litertlm (R-G, built 2026-10-27 20:14), 3507874304 B, sha256 1c14c17a7e8bd688d888300a68e2f7512b255a275a444a3f22bde395cad4b12f (the Hub tree's LFS oid at the read below); pushed to the phone and its sha256 read back against the models.tsv: \"I4B32act32rg: 3507874304 B (sha256 read back on the phone after the push) wv4b_i4b32_act32_rg.litertlm\"; every run below records model_sha256 = artifact_sha256",
            "protocol: a benchmark script, one litert_lm_advanced_main process per run, the phone-side memory guard devside_watch.sh, a ready check (thermal, CPU cap, kgsl level, rest) before each run; paths below are relative",
            "command: ./litert_lm_advanced_main --backend=gpu --model_path=/data/local/tmp/wv4b_r5/I4B32act32rg.litertlm --input_prompt_file=/data/local/tmp/wv4b_r5/prompts/bench1024.txt --max_num_tokens=2048 --max_output_tokens=256 --async=false --benchmark --benchmark_prefill_tokens=0 --benchmark_decode_tokens=0 --use_session=false --report_peak_memory_footprint=true --benchmark_prefill_tokens=1024 --benchmark_decode_tokens=256",
            "the runtime's benchmark (1,024 prompt tokens, 256 decoded, context 2048; text only, speed only); medians of 2 processes, both after the picture question's cold process had built the GPU caches (key s26|I4B32act32rg|gpu|none|text_bench warm_median)",
            "{\"tag\": \"I4B32act32rg-gpu-bA1\", \"status\": \"DONE\", \"cold\": false, \"cache_plan\": \"keep\", \"precision\": \"fp32\", \"ready_note\": \"clean; rested 48s since 15:58:11\", \"memory\": \"VmHWM_max_kB 1993288 VmSwap_max_kB 236016 MemAvailable_min_kB 2529664 kgsl_page_alloc_max_B 5342310400\", \"bench\": {\"ttft_s\": 9.15, \"prefill\": [[1024, 112.68, 9.0879]], \"decode\": [[256, 15.46, 16.5564]], \"peak_mem_mb\": 1946.57, \"marks_s\": {}}, \"init_total_ms\": 11611.41}",
            "{\"tag\": \"I4B32act32rg-gpu-bA2\", \"status\": \"DONE\", \"cold\": false, \"cache_plan\": \"keep\", \"precision\": \"fp32\", \"ready_note\": \"uncapped (thermal 1, no CPU cap, kgsl_pwrlevel 0); rested 68s since 16:02:48\", \"memory\": \"VmHWM_max_kB 1836800 VmSwap_max_kB 217960 MemAvailable_min_kB 2611908 kgsl_page_alloc_max_B 5342633984\", \"bench\": {\"ttft_s\": 9.94, \"prefill\": [[1024, 103.71, 9.8736]], \"decode\": [[256, 15.0, 17.0691]], \"peak_mem_mb\": 1793.75, \"marks_s\": {}}, \"init_total_ms\": 11476.08}"
          ],
          "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_total_ms": 11543.745,
            "max_num_tokens": 2048.0,
            "prefill_tokens": 1024.0,
            "processes": 2.0
          },
          "output_match": null,
          "peak_mem_mb": 1870.16,
          "prefill_tokens_per_s": 108.195,
          "provenance": "measured",
          "runs": true,
          "signature": null,
          "total_ops": null,
          "ttft_ms": 9545.0
        },
        "source": "data/device_runs/0.19.1/2026-11-10/wevisone-vl-4b-int4__galaxy-s26.json"
      },
      {
        "device": "mac-studio-m4-max",
        "run": {
          "accelerator": "cpu",
          "context_length": 2048,
          "date": "2026-11-12",
          "decode_tokens_per_s": 23.68,
          "delegated_ops": null,
          "env": {
            "device": "Mac Studio (M4 Max, Mac16,9)",
            "machine_label": "mac-studio-m4-max",
            "os_build": "macOS 27.0 (26A428)",
            "runtime": "litert-lm",
            "runtime_version": "0.19.1",
            "soc": "Apple M4 Max",
            "vendor_sdk": null
          },
          "error": null,
          "evidence": [
            "bundle: WeVisOne-VL-4B_int4.litertlm = the WeVisOne-VL-4B_int4.litertlm (written by hf-to-litertlm tools/externalize at 2026-11-12 08:50:35 +0900), 3507963904 B, sha256 6ae2e63e83516bf8a8f382c9938c0be8f716cba2a66b8bbcc22991e9123c1e48: the pre-upload's hash (done 2026-11-12 10:51:48) = the Hub's LFS oid at upload commit b4ed353a and at the read below = a hash of the same path after the window (2026-11-12 12:51:56 +0900 to 2026-11-12 12:53:26 +0900, the file's mtime unchanged)",
            "protocol: under the shared Mac's measurement window, one process at a time, text and picture interleaved, each process under /usr/bin/time -l and tools/fp_guard.py; both files read into memory first and held open on file descriptors, 100 % resident before and after each form (window lines below); a git-burst watcher over the window, every process CLEAN at its first attempt; runtime and machine",
            "num_cpu_threads = the runtime's default (LiteRT-LM v0.19.1 runtime/executor/llm_executor_settings.h:139 `number_of_threads = 4`): the benchmark command sets no thread count and the API call passes thread_count=None (tools/wv_rt_api.py without --cpu-thread-count); the vision encoder's thread count was not read",
            "command: /usr/bin/time -l -I tools/wv_rt_api.py --bundle WeVisOne-VL-4B_int4.litertlm --backend cpu --timing --image --prompt 'Locate all the instances that match the following description: red rectangle. blue circle. green circle.' --vision-backend cpu",
            "the picture question: r10's 512 x 512 PNG syn_a and the detection request, 297 prefill tokens and 60 decoded (the runtime's own counts); ttft_ms = from sending the question to the first streamed text, the median of 3 processes; prefill / decode = the runtime's own rates (medians); runtime_ttft_ms = the runtime's time to first token; init_time_s = engine creation; turn2_first_token_ms = the second message's first text (timed only); peak_mem_mb = the median of the processes' 'peak memory footprint' from /usr/bin/time -l in MiB (the card prints the range in GB: 17.80–17.93 GB)",
            "medians recomputed from the process records = key cpu_int4/vl; the card's row: \"| This file, CPU | 4.52 s | 5.39 s | 23.68 tok/s | 117.81 tok/s | 25.93 tok/s | 17.80–17.93 GB |\"; the manifest has no CPU row for this file (its backends list is the GPU's)",
            "First 5.3702 s, prefill 61.39, decode 23.68 tok/s, init 4.514 s, second message 0.2684 s, peak footprint 17796073712 B",
            "First 5.624 s, prefill 58.29, decode 24.25 tok/s, init 4.721 s, second message 0.2555 s, peak footprint 17929882008 B",
            "First 5.3935 s, prefill 61.20, decode 23.07 tok/s, init 4.524 s, second message 0.2695 s, peak footprint 17923246440 B",
            "answers: the three answers name the same labels in the same order (red rectangle, blue circle, green circle) with their 32 coordinates within 1 (of 1,000) of each other; the first: <ref>red rectangle</ref><box><82><460><197><601>,<197><340><370><426>,<602><162><700><269></box>, <ref>blue circle</ref><box><175><53><293><…",
            "window: [2026-11-12 12:08:37] start r4_cpu_0191 (WeVisOne-VL-4B_int4.litertlm, cpu) | 12:09:25 r4_cpu_0191 text 1 prefill 118.7 decode 26.12 rc 0 VE 0 attempt 1 CLEAN | 12:09:45 r4_cpu_0191 vl 1 first chunk 5.3702 s total 7.8093 s decode [[60, 23.67856739716848]] turn2 first 0.2684 s rc 0 VE 0 attempt 1 CLEAN | 12:10:33 r4_cpu_0191 text 2 prefill 116.73 decode 25.93 rc 0 VE 0 attempt 1 CLEAN | 12:10:54 r4_cpu_0191 vl 2 first chunk 5.624 s total 8.0145 s decode [[60, 24.245091604136558]] turn2 first 0.2555 s rc 0 VE 0 attempt 1 CLEAN | 12:11:43 r4_cpu_0191 text 3 prefill 117.81 decode 25.82 rc 0 VE 0 attempt 1 CLEAN | 12:12:03 r4_cpu_0191 vl 3 first chunk 5.3935 s total 7.8969 s decode [[60, 23.069325354328342]] turn2 first 0.2695 s rc 0 VE 0 attempt 1 CLEAN | [2026-11-12 12:12:03] end r4_cpu_0191 rc 0"
          ],
          "failure_class": null,
          "full_delegation": null,
          "latency_p50_ms": null,
          "loads": true,
          "max_abs_diff": null,
          "max_rel_diff": null,
          "metrics": {
            "decode_tokens": 60.0,
            "init_time_s": 4.524,
            "num_cpu_threads": 4.0,
            "peak_footprint_max_mib": 17099.3,
            "peak_footprint_min_mib": 16971.7,
            "prefill_tokens": 297.0,
            "processes": 3.0,
            "runtime_ttft_ms": 4896.1,
            "turn2_first_token_ms": 268.4
          },
          "output_match": null,
          "peak_mem_mb": 17092.9,
          "prefill_tokens_per_s": 61.2,
          "provenance": "measured",
          "runs": true,
          "signature": null,
          "total_ops": null,
          "ttft_ms": 5393.5
        },
        "source": "data/device_runs/0.19.1/2026-11-12/wevisone-vl-4b-int4__mac-studio-m4-max.json"
      },
      {
        "device": "mac-studio-m4-max",
        "run": {
          "accelerator": "cpu",
          "context_length": 2048,
          "date": "2026-11-12",
          "decode_tokens_per_s": 25.93,
          "delegated_ops": null,
          "env": {
            "device": "Mac Studio (M4 Max, Mac16,9)",
            "machine_label": "mac-studio-m4-max",
            "os_build": "macOS 27.0 (26A428)",
            "runtime": "litert-lm",
            "runtime_version": "0.19.1",
            "soc": "Apple M4 Max",
            "vendor_sdk": null
          },
          "error": null,
          "evidence": [
            "bundle: WeVisOne-VL-4B_int4.litertlm = the WeVisOne-VL-4B_int4.litertlm (written by hf-to-litertlm tools/externalize at 2026-11-12 08:50:35 +0900), 3507963904 B, sha256 6ae2e63e83516bf8a8f382c9938c0be8f716cba2a66b8bbcc22991e9123c1e48: the pre-upload's hash (done 2026-11-12 10:51:48) = the Hub's LFS oid at upload commit b4ed353a and at the read below = a hash of the same path after the window (2026-11-12 12:51:56 +0900 to 2026-11-12 12:53:26 +0900, the file's mtime unchanged)",
            "protocol: under the shared Mac's measurement window, one process at a time, text and picture interleaved, each process under /usr/bin/time -l and tools/fp_guard.py; both files read into memory first and held open on file descriptors, 100 % resident before and after each form (window lines below); a git-burst watcher over the window, every process CLEAN at its first attempt; runtime and machine",
            "num_cpu_threads = the runtime's default (LiteRT-LM v0.19.1 runtime/executor/llm_executor_settings.h:139 `number_of_threads = 4`): the benchmark command sets no thread count and the API call passes thread_count=None (tools/wv_rt_api.py without --cpu-thread-count); the vision encoder's thread count was not read",
            "command: /usr/bin/time -l litert-lm benchmark WeVisOne-VL-4B_int4.litertlm -p 1024 -d 256 --max-num-tokens 2048 --backend cpu --cache no --runs 1",
            "the runtime's benchmark (1,024 prompt tokens, 256 decoded, context 2,048; text only, speed only); medians of 3 processes; ttft_ms and init_time_s = the benchmark's printed time to first token and init time; peak_mem_mb = the median of the processes' 'peak memory footprint' from /usr/bin/time -l in MiB",
            "medians recomputed from the process records = key cpu_int4/text; the manifest has no CPU row for this file (its backends list is the GPU's)",
            "Prefill 118.7, decode 26.12 tok/s, init 9.152 s, first token 8.6653 s, peak footprint 16185721080 B",
            "Prefill 116.73, decode 25.93 tok/s, init 9.1112 s, first token 8.8112 s, peak footprint 16321560968 B",
            "Prefill 117.81, decode 25.82 tok/s, init 9.2014 s, first token 8.7307 s, peak footprint 16258285888 B",
            "window: [2026-11-12 12:08:37] start r4_cpu_0191 (WeVisOne-VL-4B_int4.litertlm, cpu) | 12:09:25 r4_cpu_0191 text 1 prefill 118.7 decode 26.12 rc 0 VE 0 attempt 1 CLEAN | 12:09:45 r4_cpu_0191 vl 1 first chunk 5.3702 s total 7.8093 s decode [[60, 23.67856739716848]] turn2 first 0.2684 s rc 0 VE 0 attempt 1 CLEAN | 12:10:33 r4_cpu_0191 text 2 prefill 116.73 decode 25.93 rc 0 VE 0 attempt 1 CLEAN | 12:10:54 r4_cpu_0191 vl 2 first chunk 5.624 s total 8.0145 s decode [[60, 24.245091604136558]] turn2 first 0.2555 s rc 0 VE 0 attempt 1 CLEAN | 12:11:43 r4_cpu_0191 text 3 prefill 117.81 decode 25.82 rc 0 VE 0 attempt 1 CLEAN | 12:12:03 r4_cpu_0191 vl 3 first chunk 5.3935 s total 7.8969 s decode [[60, 23.069325354328342]] turn2 first 0.2695 s rc 0 VE 0 attempt 1 CLEAN | [2026-11-12 12:12:03] end r4_cpu_0191 rc 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_time_s": 9.152,
            "max_num_tokens": 2048.0,
            "num_cpu_threads": 4.0,
            "peak_footprint_max_mib": 15565.5,
            "peak_footprint_min_mib": 15435.9,
            "prefill_tokens": 1024.0,
            "processes": 3.0
          },
          "output_match": null,
          "peak_mem_mb": 15505.1,
          "prefill_tokens_per_s": 117.81,
          "provenance": "measured",
          "runs": true,
          "signature": null,
          "total_ops": null,
          "ttft_ms": 8730.7
        },
        "source": "data/device_runs/0.19.1/2026-11-12/wevisone-vl-4b-int4__mac-studio-m4-max.json"
      },
      {
        "device": "mac-studio-m4-max",
        "run": {
          "accelerator": "gpu",
          "context_length": 2048,
          "date": "2026-10-27",
          "decode_tokens_per_s": null,
          "delegated_ops": null,
          "env": {
            "device": "Mac Studio (M4 Max, Mac16,9)",
            "machine_label": "mac-studio-m4-max",
            "os_build": "macOS 27.0 (26A428)",
            "runtime": "litert-lm",
            "runtime_version": "0.18.1",
            "soc": "Apple M4 Max",
            "vendor_sdk": null
          },
          "error": null,
          "evidence": [
            "bundle: WeVisOne-VL-4B_int4.litertlm = the wv4b_i4b32_act32_rg.litertlm (R-G, built 2026-10-27 20:14), 3507874304 B, sha256 1c14c17a7e8bd688d888300a68e2f7512b255a275a444a3f22bde395cad4b12f (the Hub tree's LFS oid at the read below); built at 20:14 (\"- **R-G** 20:14:00〜14:11(retrofit 11 s、footprint 0.41 GB)= `wv4b_i4b32_act32_rg.litertlm` 3,507,874,304 B …\"), the suite's answer files from 20:19 on; the same path read back on the phone at 22:0x and 10-28 with this sha256",
            "litert-lm 0.18.1 from PyPI; Mac Studio M4 Max (Mac16,9, macOS 27.0; the build 26A428 read 2026-11-10, the same as of 2026-10-25); paths below are relative",
            "command per question: run wv4b_i4b32_act32_rg.litertlm --prompt <question> --backend gpu --cache no --temperature 0 --seed 0 --attachment --vision-backend gpu --verbose (the answer files' cmd)",
            "reference = the source package Tencent/WeVisOne-VL @ 40f735b8 in float32 on the CPU, greedy, the same 512 x 512 pictures (arm B); output_match = the bar: image on the bf16 bar (box 0.9, count 0.6667, point 0.9, cer 0.05) and the text check (pass / fail = the float32 reference on all 10)",
            "forms.i4b32_act32_rg_gpu: {\"counts\": {\"box_matched\": 57, \"box_ref\": 58, \"count_equal\": 6, \"count_rows\": 6, \"point_matched\": 13, \"point_ref\": 14, \"exact_rows\": 3}, \"rates\": {\"box\": 0.9827586206896551, \"count\": 1.0, \"point\": 0.9285714285714286, \"cer\": 0.01639344262295082}, \"bf16\": {\"per_metric\": {\"box\": true, \"count\": true, \"point\": true, \"cer\": true}, \"pass_image\": true}, \"text_gate\": true}"
          ],
          "failure_class": null,
          "full_delegation": null,
          "latency_p50_ms": null,
          "loads": true,
          "max_abs_diff": null,
          "max_rel_diff": null,
          "metrics": {
            "box_matched": 57.0,
            "box_ref": 58.0,
            "count_equal": 6.0,
            "count_rows": 6.0,
            "ocr_cer": 0.01639344262295082,
            "point_matched": 13.0,
            "point_ref": 14.0,
            "rows_token_identical": 3.0,
            "text_pass_fail_same": 10.0,
            "text_rows": 10.0,
            "text_same_as_reference": 6.0
          },
          "output_match": true,
          "peak_mem_mb": null,
          "prefill_tokens_per_s": null,
          "provenance": "measured",
          "runs": true,
          "signature": "the suite: 29 picture questions (17 pictures, 512 x 512 PNG) + 10 text questions, one process per question, greedy",
          "total_ops": null,
          "ttft_ms": null
        },
        "source": "data/device_runs/0.18.1/2026-10-27/wevisone-vl-4b-int4__mac-studio-m4-max.json"
      },
      {
        "device": "mac-studio-m4-max",
        "run": {
          "accelerator": "gpu",
          "context_length": 2048,
          "date": "2026-11-12",
          "decode_tokens_per_s": 92.34,
          "delegated_ops": null,
          "env": {
            "device": "Mac Studio (M4 Max, Mac16,9)",
            "machine_label": "mac-studio-m4-max",
            "os_build": "macOS 27.0 (26A428)",
            "runtime": "litert-lm",
            "runtime_version": "0.19.1",
            "soc": "Apple M4 Max",
            "vendor_sdk": null
          },
          "error": null,
          "evidence": [
            "bundle: WeVisOne-VL-4B_int4.litertlm = the WeVisOne-VL-4B_int4.litertlm (written by hf-to-litertlm tools/externalize at 2026-11-12 08:50:35 +0900), 3507963904 B, sha256 6ae2e63e83516bf8a8f382c9938c0be8f716cba2a66b8bbcc22991e9123c1e48: the pre-upload's hash (done 2026-11-12 10:51:48) = the Hub's LFS oid at upload commit b4ed353a and at the read below = a hash of the same path after the window (2026-11-12 12:51:56 +0900 to 2026-11-12 12:53:26 +0900, the file's mtime unchanged)",
            "protocol: under the shared Mac's measurement window, one process at a time, text and picture interleaved, each process under /usr/bin/time -l and tools/fp_guard.py; both files read into memory first and held open on file descriptors, 100 % resident before and after each form (window lines below); a git-burst watcher over the window, every process CLEAN at its first attempt; runtime and machine",
            "command: /usr/bin/time -l -I tools/wv_rt_api.py --bundle WeVisOne-VL-4B_int4.litertlm --backend gpu --timing --image --prompt 'Locate all the instances that match the following description: red rectangle. blue circle. green circle.' --vision-backend gpu",
            "the picture question: r10's 512 x 512 PNG syn_a and the detection request, 297 prefill tokens and 60 decoded (the runtime's own counts); ttft_ms = from sending the question to the first streamed text, the median of 3 processes; prefill / decode = the runtime's own rates (medians); runtime_ttft_ms = the runtime's time to first token; init_time_s = engine creation; turn2_first_token_ms = the second message's first text (timed only); peak_mem_mb = the median of the processes' 'peak memory footprint' from /usr/bin/time -l in MiB (the card prints the range in GB: 7.13–7.14 GB)",
            "medians recomputed from the process records = key gpu/vl; the card's row: \"| This file, GPU | 2.52 s | 0.49 s | 92.34 tok/s | 1,249.49 tok/s | 89.17 tok/s | 7.13–7.14 GB |\"; the manifest's row equals these processes' ranges",
            "First 0.4835 s, prefill 877.31, decode 92.92 tok/s, init 2.508 s, second message 0.1182 s, peak footprint 7128764520 B",
            "First 0.4851 s, prefill 876.47, decode 92.34 tok/s, init 2.527 s, second message 0.1183 s, peak footprint 7144476752 B",
            "First 0.4858 s, prefill 875.19, decode 91.89 tok/s, init 2.52 s, second message 0.1184 s, peak footprint 7141249104 B",
            "answers: the three answers are the same text; the first: <ref>red rectangle</ref><box><82><460><197><601>,<197><340><370><427>,<602><162><700><269></box>, <ref>blue circle</ref><box><175><53><292><…",
            "window: [2026-11-12 12:07:16] start r4_gpu_0191 (WeVisOne-VL-4B_int4.litertlm, gpu) | 12:07:33 r4_gpu_0191 text 1 prefill 1252.27 decode 89.17 rc 0 VE 0 attempt 1 CLEAN | 12:07:44 r4_gpu_0191 vl 1 first chunk 0.4835 s total 1.1043 s decode [[60, 92.91736718231253]] turn2 first 0.1182 s rc 0 VE 0 attempt 1 CLEAN | 12:08:00 r4_gpu_0191 text 2 prefill 1248.3 decode 89.17 rc 0 VE 0 attempt 1 CLEAN | 12:08:10 r4_gpu_0191 vl 2 first chunk 0.4851 s total 1.109 s decode [[60, 92.33673107316275]] turn2 first 0.1183 s rc 0 VE 0 attempt 1 CLEAN | 12:08:26 r4_gpu_0191 text 3 prefill 1249.49 decode 88.97 rc 0 VE 0 attempt 1 CLEAN | 12:08:36 r4_gpu_0191 vl 3 first chunk 0.4858 s total 1.1113 s decode [[60, 91.88722583372967]] turn2 first 0.1184 s rc 0 VE 0 attempt 1 CLEAN | [2026-11-12 12:08:36] end r4_gpu_0191 rc 0"
          ],
          "failure_class": null,
          "full_delegation": null,
          "latency_p50_ms": null,
          "loads": true,
          "max_abs_diff": null,
          "max_rel_diff": null,
          "metrics": {
            "decode_tokens": 60.0,
            "init_time_s": 2.52,
            "peak_footprint_max_mib": 6813.5,
            "peak_footprint_min_mib": 6798.5,
            "prefill_tokens": 297.0,
            "processes": 3.0,
            "runtime_ttft_ms": 349.7,
            "turn2_first_token_ms": 118.3
          },
          "output_match": null,
          "peak_mem_mb": 6810.4,
          "prefill_tokens_per_s": 876.47,
          "provenance": "measured",
          "runs": true,
          "signature": null,
          "total_ops": null,
          "ttft_ms": 485.1
        },
        "source": "data/device_runs/0.19.1/2026-11-12/wevisone-vl-4b-int4__mac-studio-m4-max.json"
      },
      {
        "device": "mac-studio-m4-max",
        "run": {
          "accelerator": "gpu",
          "context_length": 2048,
          "date": "2026-11-12",
          "decode_tokens_per_s": 89.17,
          "delegated_ops": null,
          "env": {
            "device": "Mac Studio (M4 Max, Mac16,9)",
            "machine_label": "mac-studio-m4-max",
            "os_build": "macOS 27.0 (26A428)",
            "runtime": "litert-lm",
            "runtime_version": "0.19.1",
            "soc": "Apple M4 Max",
            "vendor_sdk": null
          },
          "error": null,
          "evidence": [
            "bundle: WeVisOne-VL-4B_int4.litertlm = the WeVisOne-VL-4B_int4.litertlm (written by hf-to-litertlm tools/externalize at 2026-11-12 08:50:35 +0900), 3507963904 B, sha256 6ae2e63e83516bf8a8f382c9938c0be8f716cba2a66b8bbcc22991e9123c1e48: the pre-upload's hash (done 2026-11-12 10:51:48) = the Hub's LFS oid at upload commit b4ed353a and at the read below = a hash of the same path after the window (2026-11-12 12:51:56 +0900 to 2026-11-12 12:53:26 +0900, the file's mtime unchanged)",
            "protocol: under the shared Mac's measurement window, one process at a time, text and picture interleaved, each process under /usr/bin/time -l and tools/fp_guard.py; both files read into memory first and held open on file descriptors, 100 % resident before and after each form (window lines below); a git-burst watcher over the window, every process CLEAN at its first attempt; runtime and machine",
            "command: /usr/bin/time -l litert-lm benchmark WeVisOne-VL-4B_int4.litertlm -p 1024 -d 256 --max-num-tokens 2048 --backend gpu --cache no --runs 1",
            "the runtime's benchmark (1,024 prompt tokens, 256 decoded, context 2,048; text only, speed only); medians of 3 processes; ttft_ms and init_time_s = the benchmark's printed time to first token and init time; peak_mem_mb = the median of the processes' 'peak memory footprint' from /usr/bin/time -l in MiB",
            "medians recomputed from the process records = key gpu/text; the manifest's row equals these processes' ranges",
            "Prefill 1252.27, decode 89.17 tok/s, init 5.3198 s, first token 0.8289 s, peak footprint 5331979408 B",
            "Prefill 1248.3, decode 89.17 tok/s, init 5.3569 s, first token 0.8315 s, peak footprint 5349575848 B",
            "Prefill 1249.49, decode 88.97 tok/s, init 5.1393 s, first token 0.8308 s, peak footprint 5339401384 B",
            "window: [2026-11-12 12:07:16] start r4_gpu_0191 (WeVisOne-VL-4B_int4.litertlm, gpu) | 12:07:33 r4_gpu_0191 text 1 prefill 1252.27 decode 89.17 rc 0 VE 0 attempt 1 CLEAN | 12:07:44 r4_gpu_0191 vl 1 first chunk 0.4835 s total 1.1043 s decode [[60, 92.91736718231253]] turn2 first 0.1182 s rc 0 VE 0 attempt 1 CLEAN | 12:08:00 r4_gpu_0191 text 2 prefill 1248.3 decode 89.17 rc 0 VE 0 attempt 1 CLEAN | 12:08:10 r4_gpu_0191 vl 2 first chunk 0.4851 s total 1.109 s decode [[60, 92.33673107316275]] turn2 first 0.1183 s rc 0 VE 0 attempt 1 CLEAN | 12:08:26 r4_gpu_0191 text 3 prefill 1249.49 decode 88.97 rc 0 VE 0 attempt 1 CLEAN | 12:08:36 r4_gpu_0191 vl 3 first chunk 0.4858 s total 1.1113 s decode [[60, 91.88722583372967]] turn2 first 0.1184 s rc 0 VE 0 attempt 1 CLEAN | [2026-11-12 12:08:36] end r4_gpu_0191 rc 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_time_s": 5.32,
            "max_num_tokens": 2048.0,
            "peak_footprint_max_mib": 5101.8,
            "peak_footprint_min_mib": 5085.0,
            "prefill_tokens": 1024.0,
            "processes": 3.0
          },
          "output_match": null,
          "peak_mem_mb": 5092.0,
          "prefill_tokens_per_s": 1249.49,
          "provenance": "measured",
          "runs": true,
          "signature": null,
          "total_ops": null,
          "ttft_ms": 830.8
        },
        "source": "data/device_runs/0.19.1/2026-11-12/wevisone-vl-4b-int4__mac-studio-m4-max.json"
      }
    ]
  },
  "model": {
    "family": "wevisone-vl",
    "id": "wevisone-vl-4b-int4",
    "license": "apache-2.0",
    "source_url": "https://huggingface.co/litert-community/WeVisOne-VL-4B",
    "task": "image-text-to-text"
  },
  "pitfalls": [
    "External weights (since 2026-11-12): on an Android GPU the file needs LiteRT-LM 0.19.0 or later; the GPU engine of the v0.18.0 and v0.18.1 Android builds crashes (SIGSEGV) while it is being built (LiteRT-LM#3986). For those builds the earlier file at revision ee03d010 keeps every weight inside the graph; the new file holds the same weight bytes and gave the same answers as the earlier one to 15 questions on the Galaxy S26 GPU (LiteRT-LM 0.19.1) and to 3 on the Mac GPU (L33, L164).",
    "Which bytes the numbers are: the card's Mac table ran the file of 2026-11-12; every other measurement on the card (2026-10-27 to 2026-11-11) ran the earlier file, and on the phone speed, engine creation and memory were not measured again (L164).",
    "One picture size: the file always reads 512 x 512, and every comparison on the card is at that size; turn the picture upright from its EXIF orientation and stretch it with Pillow's BICUBIC filter, no padding or cropping, saved as PNG (other resizing not tested). The source package chooses the resolution per picture; how far the original's answers at its own resolution differ was not measured for the 4B (L37, L143).",
    "Text-only conversations: without a picture, every token after the 271st of the conversation (prompt, answer and chat template together) gets a wrong position, and from there the answer can differ from the original's; conversations with a picture are not affected (L144).",
    "Later messages: from the second message on, LiteRT-LM passes the previous answer's stop token to the model with a zero embedding, so the second answer starts from a context the original never sees; ask one question per conversation and send the picture again for the next (L73, L145).",
    "Two LiteRT-LM 0.19.1 builds for Android: the tag build links the GPU accelerator in from source. On the Galaxy S26, with this file, it prefilled the 1,024-token text benchmark faster (108.20 against 90.45 tok/s) but reached the picture question's first token later (4.86 against 4.22 s) than the same commit loading the prebuilt libLiteRtGpuAccelerator.so, in both pairs of an A B B A session; only the tag's default build had its answers checked (L79-L84, L165).",
    "Segmentation and matting: the file does not contain the source model's 9.6M-parameter pixel decoder; to those templates it answers with a box that ends in an <OBJ> token, and no mask (L26, L69).",
    "Absent objects: to a referring question with no matching object it answered <box>none</box>; asked to count airplanes with boxes in a radar picture that holds none, it answered one small box and 'Total number: 1.' — the original does the same both times (L69).",
    "Startup on the Galaxy S26 (build A): engine creation 11.5 s, then opening the picture conversation another 4.9 s, then the first token at 4.86 s; the first run writes about 0.46 GB of cache files next to the model (L50, L81, L84).",
    "Memory: on LiteRT-LM 0.19.1 the 12 GB Galaxy S26 ran 57 processes of this file and stopped none; the highest peak memory of any process was 4.02 GB, the picture question's 3.68-3.80 GB over the two builds; phones with less memory were not tried (L81-L82, L146).",
    "A CPU: use the repo's int8 file; this file's answers were not checked on a CPU. On a Mac CPU (Apple M4 Max, the PyPI litert-lm 0.19.1 package, the decoder with the runtime's default 4 threads, no cache files) it ran only speed rows: the picture's first token 5.39 s, decode 23.68 tok/s, peak memory 17.80-17.93 GB, against 0.49 s, 92.34 tok/s and 7.13-7.14 GB on that Mac's GPU (L33, L103-L111, L147).",
    "Android program: the v0.19.1 release has no Android binary; the card's runs use litert_lm_advanced_main built from the tag with the tag's prebuilt/android_arm64 libraries beside it and LD_LIBRARY_PATH set to them, one process per answer, the prompt file holding [image:<path>] and then the question with nothing between them and no newline at the end (L44-L47).",
    "Sampling: the file carries no sampler settings, so LiteRT-LM 0.19.1 decodes it greedily; the source package's repetition penalty of 1.05 was not tested (L50, L147).",
    "Not tested: the known-answer and agreement tests on a CPU or on a desktop with LiteRT-LM 0.19.1, the files of 2026-11-12 in an app (the card's one test app ran the earlier file) or on Android with LiteRT-LM 0.18.x, other apps, sampled answers; there is no NPU build (L54, L147)."
  ],
  "schema_version": "1.2"
}
