A very cheap multimodal agent model, and the weights are out
Hy4 Flash costs $0.07 per million input tokens. The 441 GB weights are for big rigs only.
Hy Flash read an invoice image and wrote the right row into SQLite in about 30 seconds, flagging oddities instead of guessing. It is Tencent's cheap multimodal model, and this week it published its weights.
The deeper look
Plotly tested it on a data-analytics bench and reports it beat GPT Luna while costing $0.13 for 43 questions. That is a vendor-run benchmark from a third party with a product stake, so read it as encouraging, not settled.
The hands-on evidence is Fahd Mirza's, and it is mixed in a useful way. The invoice job took 30 seconds and a three-minute multilingual cafe page rendered Chinese, Hindi and Arabic properly. But sorting nine messy files by computer use took about 13 minutes, though correctly.
Sources disagree on size. Tencent says 12B active; a Hacker News comment said about 64B, likely mixing in the 64B n-gram embedding table that ships with the weights. Trust Tencent's model card here.
It is for builders who run many cheap calls over documents and screenshots. Skip it for latency-sensitive work, and do not plan on running it locally without a multi-GPU server.
How to start: the API docs are at hy.tencent.ai/research/hy4-flash. Weights: huggingface.co/tencent/Hy4-flash. Community quantizations exist if you have the memory.
$ ollama run hf.co/AngelSlim/Hy4-flash-GGUF$ llama-server -hf AngelSlim/Hy4-flash-GGUF$ vllm serve tencent/Hy4-flash$ hf download tencent/Hy4-flash$ curl https://openrouter.ai/api/v1/chat/completions -H "Authorization: Bearer $OPENROUTER_API_KEY" -H "Content-Type: application/json" -d '{"model": "tencent/hy4-flash", "messages": [{"role": "user", "content": "Hello"}]}'