infosave / Hy-MT2-cmf

Tencent Hy-MT2 — CMF (1.8B · 7B · 30B-A3B)

Tencent's Hy-MT2 translation family — 33 languages, instruction-following translation (terminology, style, placeholders, structured data) — as CMF files for cortiq : one memory-mapped file per model with the weights, the vendor tokenizer, the exact chat template and per-tensor hashes. The same file runs on the CPU, on Vulkan (NVIDIA / AMD / Intel), DX12 and Apple Metal; no Python, no PyTorch, no CUDA toolkit.

cargo install cortiq-cli --version 0.7.0        # first release with HunYuan support

hf download infosave/Hy-MT2-cmf hy-mt2-7b-q4tp.cmf --local-dir .
cortiq verify hy-mt2-7b-q4tp.cmf
cortiq run hy-mt2-7b-q4tp.cmf --greedy --prompt "Translate the following text into English. Note that you should only output the translated result without any additional explanation:

Сегодня хорошая погода, и мы пойдём гулять в парк."

Requires cortiq 0.7.0 or newer. Earlier runtimes do not know the hunyuan_v1_dense / hy_v3 architectures (and 0.6.9 crashes on any 4-bit embedding — fixed in the same release).

Files

¹ Twelve 512-token windows of wikitext-2 test, exact attention, the same yardstick every card on this account uses. The bf16-class reference for the 1.8B ( q8_2f , not published) scores 24.77 — the 4-bit file sits within 10% (27.13 vs 24.77) of it. These are translation models: their perplexity on English encyclopedia prose is not what they were trained for, which is why column ² exists.

² Sixteen chat-formatted sentence pairs (RU/EN/ZH/DE/FR/ES/JA/IT/UK/PT, 1 073 tokens) in the model's own prompt format — the number that tracks what the file is for. The q1t file is Tencent's own 1.25-bit quantization-aware checkpoint carried over exactly (see below); it is a translation-only specialist and its general-text perplexity is not meaningful, but its translations are — see the samples further down.

³ Not a conversion defect: llama.cpp with the STQ1_0 kernel (PR #22836) scores the very same GGUF at 1 793 on its own wikitext-2 protocol (first twelve 512-token chunks, BOS-anchored, second halves scored), our cold-window protocol at 4 460 — both say the 1.25-bit checkpoint no longer models free English prose, while it translates correctly on every backend. On the same Xeon that llama.cpp build decodes it at 2.2 tok/s; cortiq's q1t kernels at 21.9.

⁴ The 7B uses HunYuan's older tokenizer (vocab 128 167, GPT-4-style splitting, <|startoftext|> … <|extra_0|> chat markup) and is sensitive to running without its BOS in the middle of a corpus; its translations are the best of the three dense files. Compare sizes on column ², not here.

Every q4tp file was quantized straight from the bf16 safetensors, streamed shard by shard; nothing here is a re-quantization of a lower-precision file. The tied lm_head of the dense models is not written twice: the embedding serves both ends, as in the source.

What the runtime had to learn

Two things distinguish HunYuan from the Qwen/Llama block, and both ride in the file header rather than in flags:

• Dense 1.8B / 7B ( hunyuan_v1_dense ) apply the per-head q/k RMSNorm after RoPE ( query_layernorm / key_layernorm on the rotated vectors). That is not a reordering you can fold into the weights — the rotation preserves a head's norm but not its elementwise norm weights — so the engine carries an explicit qk_norm_after_rope flag through the CPU, Metal and WGSL rope kernels. Their "dynamic" NTK-alpha RoPE ( alpha = 1000 ) is one rescaled base, 10000 · 1000^(128/126) = 11 158 840 , applied at conversion; the full 262 144-token window is native, nothing is rescaled per position.
• 30B-A3B ( hy_v3 ) is 48 layers with layer 0 dense (6 912 wide) and 47 sparse layers of 128 experts, top-8 by sigmoid score plus a selection bias ( expert_bias , DeepSeek-V3 style: bias for the choice, unbiased scores for the weights), renormalized and multiplied by router_scaling_factor = 2.826 , plus one always-on shared expert of
  1. The converter maps mlp.router.gate / mlp.shared_mlp onto the canonical layout and the header carries the router constants; the wgpu whole-token graph learned the two pieces it lacked — the routed scale and a shared expert without a gate — so the 30B decodes on the card as one graph (12 submits per token) instead of falling to the per-op path (145 submits, 1.2 tok/s) the way every sigmoid-scaled MoE did before 0.7.0.

The 1.25-bit file, exactly

Tencent ships Hy-MT2-1.8B-1.25Bit-GGUF : a quantization-aware ternary checkpoint in llama.cpp's STQ1_0 type (256-weight blocks, one f16 scale, and in every group of four lanes exactly one zero and three ±1 — 5 bits per 4 weights). cortiq import-gguf decodes that block layout natively and re-encodes each 32-group as CMF q1t (ternary, base-3 packed, f16 scale, empty outlier overlay): the reconstruction is bit-for-bit the values the GGUF stores — no calibration, no second quantizer. The cost of the general-purpose container is size: 2.25 bits/weight against 1.31, so the file is 695 MB against 462, with the token table at q8_2f (Tencent stores it at 6.5 bits).

cortiq import-gguf Hy-MT2-1.8B-1.25Bit.gguf --quant q1t \
    --tokenizer-dir ./Hy-MT2-1.8B --output hy-mt2-1.8b-q1t.cmf

--tokenizer-dir matters: HunYuan's pre-tokenizer splits digits in runs of one to three and isolates CJK before the byte-level step, which a tokenizer rebuilt from GGUF metadata cannot express. The flag embeds the vendor tokenizer.json and chat template verbatim, so the ternary file tokenizes exactly like the 4-bit ones.

Measured

Steady-state decode, single stream, cortiq bench --core --tokens 128 --ignore-eos , cortiq 0.7.0. The dense files are latency-bound on a discrete card (a 1 GB model needs ~8 submits per token), so the CPU matters as much as the GPU there; the MoE row is where the card counts.

The MoE on any card (dynamic loading)

The 30B's whole-token graph is not all-or-nothing: as many leading layers as the VRAM budget admits stay resident on the card, the host finishes the rest — one boundary crossing per token, no expert paging. Measured on the RTX PRO 4000 with CMF_GPU_VRAM_MB capped to what each card size would auto-detect (the CPU alone: 10.5 tok/s on this 14-core Xeon):

Perplexity and the greedy continuation do not move along the ladder — the split changes where a layer runs, never what it computes. CMF_GPU_VRAM_MB overrides the auto-detected budget when you want to cap it by hand.

⁵ Not measured: the 15.8 GB file is at the edge of a 24 GB Mac (the 14.3 GB Qwen3.8-27B decodes at 5.7 tok/s there) and on Metal a sigmoid-routed, ungated-shared MoE layer still runs its experts on the CPU — the Metal select kernel is the next port.

Prompt ingest (41-token prompt): on the RTX PRO 4000 the dense files take 180 (1.8B) and 118 (7B) tok/s; the 30B ingests at 8 tok/s on this card — its batched prefill re-stages the expert buffers per 32-token chunk instead of sharing the decode graph's resident copy, so a long source paragraph costs seconds before the first token. Decode is unaffected; sharing the buffers is the next item on the MoE list. On the M4: 469 tok/s for the 1.8B q4tp, 218 for the ternary file, 122 for the 7B.

First answer vs. the rest. On a discrete card the weights are uploaded when the whole-token graph is first built — 18.5 GB for the 30B, ~26 s on this box — and cortiq run folds that into its printed "decode" figure for a one-shot prompt, which then reads a few tok/s. The per-token cost after the upload is the table above ( cortiq bench --core measures it after an untimed warm-up). cortiq serve pays the upload once per process, so every request after the first decodes at the steady rate.

Prompting

The models have no system prompt. The chat template is embedded, so cortiq run and cortiq serve wrap a plain user message correctly; what goes into the message is Tencent's instruction, with the language name written out in the prompt's language:

Translate the following text into {target_lang}. Note that you should only output the translated result without any additional explanation:

{source_text}

将以下文本翻译为{target_lang}，注意只需要输出翻译后的结果，不要额外解释：

{source_text}

Terminology, style, personalization, delimiter-preserving and structured-data prompts are documented on the source card . Recommended sampling (Tencent): 1.8B / 7B — temperature 0.7, top-p 0.6, top-k 20, repetition penalty 1.05; 30B-A3B — temperature 0.7, top-p 1.0, no repetition penalty. --greedy is the deterministic choice for evaluation.

Samples from the 1.8B files, greedy, identical on CPU and GPU:

Server

cortiq serve hy-mt2-7b-q4tp.cmf --port 8080
curl http://localhost:8080/v1/chat/completions -H 'Content-Type: application/json' \
  -d '{"model":"hy-mt2-7b-q4tp","temperature":0,"messages":[{"role":"user","content":"Translate the following text into French. Note that you should only output the translated result without any additional explanation:\n\nThe meeting is at ten."}]}'

/v1/chat/completions , /v1/completions and /v1/models are live; --ollama adds an Ollama-compatible listener.

Reproduce

cortiq convert --model tencent/Hy-MT2-1.8B    --quant q4tp --output hy-mt2-1.8b-q4tp.cmf
cortiq convert --model tencent/Hy-MT2-7B      --quant q4tp --output hy-mt2-7b-q4tp.cmf
cortiq convert --model tencent/Hy-MT2-30B-A3B --quant q4tp --output hy-mt2-30b-a3b-q4tp.cmf

Streaming from the hub: peak disk is the output file. Every file was size- and hash-verified after upload ( cortiq verify ).

Weights derive from Tencent's release and remain under its Apache-2.0 terms. The CMF container and the cortiq runtime are Apache-2.0 as well (see the repository's LICENSE and PATENTS.md).