DeepSeek V4.1 Flash vs Hy3

DeepSeek V4.1 Flash

A large multimodal agent model with a 1M-token context, documented tool calls, JSON Schema output and a reasoning budget, under MIT. It needs about 612 GB of memory across several GPUs in vLLM or SGLang. The main risks are prompts built from DeepSeek's code rather than a template, no published safety results and no answers in its repositories.

Hy3

Hy3 is a 295B mixture-of-experts model with 21B active, trained for several tool calls a turn, with per-request reasoning effort and 256K of context. It needs an 8-GPU server for FP8 or about 200 GB for a 4-bit build. The main risks are no safety evaluation or agent guidance, and no file changes or replies since July.

DeepSeek V4.1 Flash

1. DeepSeek's reference run is its inference folder, convert.py with --model-parallel 8 --expert-dtype fp4, then torchrun --nproc-per-node 8 generate.py --ckpt-path "$SAVE_PATH" --config config.json --interactive
2. In vLLM 0.32.0 the parsers are named deepseek_v41 for both --tool-call-parser and --reasoning-parser. That's vLLM's naming, DeepSeek doesn't document a vLLM command
3. Build prompts with encoding.py or pip install deepseek-recipe, not the September repository's chat_template.jinja, which an open pull request says trails the 16 October encoding
4. Tool calls come back in a DSML calls block with a space before calls, invoke and parameter, unlike V4
5. Set reasoning_effort to 1 to 100, or low, high or max (50, 75, 100), in thinking mode only. Default is 75
6. Sample at temperature 1.0 and top_p 1.0 and allow 256K or more output tokens, as the card recommends
7. Without the hardware, call DeepSeek's API as deepseek-flash, which serves this exact update

Hy3

1. Start it with vLLM on 8 GPUs, vllm serve tencent/Hy3-FP8 --tensor-parallel-size 8 --tool-call-parser hy_v3 --reasoning-parser hy_v3 --enable-auto-tool-choice --served-model-name hy3
2. In SGLang use --tool-call-parser hunyuan --reasoning-parser hunyuan, as the card shows
3. Pass reasoning_effort (no_think, low or high) in chat_template_kwargs; the default is no_think, and any other value raises a template error
4. With the vLLM reasoning parser the chain of thought comes back in message.reasoning and the answer in message.content
5. Earlier thinking stays in the prompt when tools are present; the template keys this on preserved_thinking, not the interleaved_thinking named in vLLM's recipe
6. Use temperature 0.9 and top_p 1.0, the card's recommended sampling
7. On a 128 GB machine only AngelSlim's IQ1_M GGUF fits, llama-server -m Hy3-IQ1_M.gguf -ctk q8_0 -ctv q8_0 -fa on -c 65536, with llama.cpp b11768 or later; at 89.4 GB × 1.2 it sits right at the 107.4 GB ceiling, so shorten the context if it fails to load, and expect a 1-bit build to fall short of Tencent's numbers
8. Without the hardware, call it through Featherless AI or DeepInfra via Hugging Face's inference providers