{ "manifest_schema": "0.1.2", "repo": "litert-community/Hy-MT2-1.8B", "generated": "2026-09-22", "generator": "make_manifest.py", "model": { "display_name": "Hy-MT2-1.8B", "base_model": "tencent/Hy-MT2-1.8B", "architecture": "Dense (HunYuanDenseV1ForCausalLM, 32L, 2.04B): GQA attention with QK-norm, hidden 2048, 16 query / 4 KV heads, head_dim 128, intermediate 6144, vocab 120818, tied embeddings. Multilingual translation model (33 languages, instruction-following translation, not open chat)", "parameters_b": 2.04, "license": "apache-2.0", "context_length": 4096, "capabilities": { "vision": false, "audio": false, "thinking": { "declared": false } } }, "variants": [ { "file": "Hy-MT2-1.8B_int8.litertlm", "sha256": "529e6d378df5869d89a5a08717c06604105a32d8a4dab4800175d6baabc4da50", "size_bytes": 1815622960, "sections": [ { "type": "LlmMetadataProto", "size_bytes": 1056 }, { "type": "HF_Tokenizer_Zlib", "size_bytes": 2006574 }, { "type": "TFLiteModel", "size_bytes": 1813574960, "model_type": "tf_lite_prefill_decode" } ], "quantization": "export-time dynamic int8 (linears + embedding)", "backends": [ "cpu", "gpu" ], "default_backend": "cpu", "recommended": [ { "platform": "android", "device_class": "flagship", "backend": "gpu", "reason": "measured on one Galaxy S26 with one binary (S7 Android backfill 2026-09-05, 205-token prompt, 2 cold runs per backend, cooled before each run): gpu decodes 20.38-20.84 tok/s against cpu's 12.27-16.97 (1.4x), prefill 385.7 vs 126.3-217.2 tok/s, peak 742 vs 2760 MB, engine init 3.5 vs 6.0 s. Same-device CPU control taken - a measured win for gpu on decode. Gate prompt was the card's translation instruction (English to Japanese); both backends returned a correct translation" }, { "platform": "android", "device_class": "midrange", "backend": "cpu", "reason": "measured on one Pixel 8a with one binary (S7 Android backfill 2026-09-05, 2 cold runs, 205-token prompt, cooled before each run): cpu 46.1-50.8 tok/s prefill / 6.37-8.3 decode, peak 2685 MB, engine init 14.3 s. The GPU path was not attempted on this 8 GB phone: a 1.82 GB int8 file with the Mali GPU's roughly 4x file-size residency would exceed its memory budget, so this is a verified choice, not a measured win. Gate prompt was the card's translation instruction (English to Japanese) and the CPU returned a correct translation. On the Galaxy S26 the flagship recommendation is gpu (1.4x)" } ], "requirements": { "platform_notes": [ "Measured with litert-lm 0.16.0; the bundle is a plain dense prefill/decode export (generic_model)", "Translation model: prompt with the source card's translation instructions (e.g. 'Translate the following text into . Note that you should **only output the translated result without any additional explanation**:'), not open chat", "Metadata start token deliberately dropped: the chat template renders <|hy_begin_of_sentence|> itself and the engine prepends the metadata start_token unconditionally (proved inside the runtime: [start_token]+prompt and [template-BOS]+prompt generate byte-identical greedy output), so the default export fed BOS twice. This file carries the training stream", "Source rope_scaling {type: dynamic, alpha: 1000} is baked statically at conversion (rope_theta 11158839.925); inv_freq and teacher-forced logits are bitwise-equal to the HF reference, valid to 262144 positions - far past this bundle's 4096 context" ] }, "measured": [ { "device": "Apple M4 Max", "os": "macOS", "backend": "cpu", "runtime": "litert-lm 0.16.0", "prompt_tokens": 256, "decode_tokens": 256, "prefill_tps": 210.95, "decode_tps": 33.97, "ttft_s": 1.2431, "max_num_tokens": 4096, "runs": 3, "date": "2026-08-27", "source": "litertlm-convert reports/shiplogs_20260827/bench_hymt2_cpu.log; litert-lm benchmark -p 256 -d 256 --runs 3 --cache no. Second protocol run: prefill 209.98 / decode 32.37 / ttft 1.2500 (prefill spread ~0.5%, decode ~5%). Quiet host (load ~3.3)" }, { "device": "Apple M4 Max", "os": "macOS", "backend": "gpu", "runtime": "litert-lm 0.16.0", "prompt_tokens": 256, "decode_tokens": 256, "prefill_tps": 2008.37, "decode_tps": 105.8, "ttft_s": 0.1369, "max_num_tokens": 4096, "runs": 3, "date": "2026-08-27", "source": "litertlm-convert reports/shiplogs_20260827/bench_hymt2_gpu.log; litert-lm benchmark -p 256 -d 256 --runs 3 --cache no, taken after a >=300 s rest. Backend gated on real generations first (8-question gate 6/8 via litert-lm run --backend gpu, same misses as CPU). Second protocol run: prefill 2002.78 / decode 105.39 / ttft 0.1373 (~0.3% spread)" }, { "device": "Galaxy S26 (SM-S942Q, Qualcomm SM8850, Adreno)", "backend": "gpu", "runtime": "litert_lm_advanced_main v0.16.0, LiteRT CL delegate", "prompt_tokens": 197, "decode_tokens": 93, "prefill_tps": 385.7, "decode_tps": "20.38-20.84", "ttft_s": 0.56, "load_s": 3.5, "peak_memory_mb": 742, "cache": "no", "runs": 2, "date": "2026-09-05", "source": "S7 Android backfill: 205-token prompt with --benchmark, 2 cold runs (compile/weight caches deleted between runs, device cooled below 42 C before each run); GPU gate full LITERT_CL delegation (2848 ops across 2 subgraphs), generation verified on the gate prompt; the cpu row of the same date is the same-device control" }, { "device": "Galaxy S26 (SM-S942Q, Qualcomm SM8850, Adreno)", "backend": "cpu", "runtime": "litert_lm_advanced_main v0.16.0 (XNNPACK CPU)", "prompt_tokens": 197, "decode_tokens": 88, "prefill_tps": "126.3-217.2", "decode_tps": "12.27-16.97", "ttft_s": "0.97-1.64", "load_s": 6.0, "peak_memory_mb": 2760, "cache": "no", "runs": 2, "date": "2026-09-05", "source": "S7 Android backfill: same-device CPU control - identical binary, prompt file and flags, only --backend differs; 205-token prompt with --benchmark, 2 cold runs (compile/weight caches deleted between runs, device cooled below 42 C before each run); same-device control for the gpu row of the same date" }, { "device": "Pixel 8a (Tensor G3, Mali-G715, 7.75 GB RAM)", "backend": "cpu", "runtime": "litert_lm_advanced_main v0.16.0 (XNNPACK CPU)", "prompt_tokens": 197, "decode_tokens": 88, "prefill_tps": "46.1-50.8", "decode_tps": "6.37-8.3", "ttft_s": "4.0-4.43", "load_s": 14.3, "peak_memory_mb": 2685, "cache": "no", "runs": 2, "date": "2026-09-05", "source": "S7 Android backfill: same-device CPU control - identical binary, prompt file and flags, only --backend differs; 205-token prompt with --benchmark, 2 cold runs (compile/weight caches deleted between runs, device cooled below 42 C before each run); the GPU path was not run for this file on this handset (1.82 GB int8: Mali GPU residency ~4x file size - not attempted), so this row has no GPU counterpart" } ], "known_issues": [ "8-question sanity gate scores 6/8 on BOTH backends with the same two misses ('Cool' for opposite-of-hot, 'pink' for the rhyme) - a property of this translation-tuned 1.8B, not of a backend. Arithmetic, factual and translation items are all correct; no degeneration", "Translation greedy A/B vs HF bf16 (source README default-translation prompt): byte-identical on 1 of 3 probes, fluent int8-class alternates on the other two (e.g. spectaculaire -> significative)" ] } ] }