Hy3 vs LFM2.5-2.6B

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.

LFM2.5-2.6B

A small, fast cousin for tool calls, extraction and long-context steps on a laptop or phone. Liquid's own 4-bit build is 1.6 GB, and every common runtime runs it. It always thinks, isn't meant for coding or knowledge-heavy work, and organisations over $10M of revenue need another licence. There are no safety results or security contact.

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

LFM2.5-2.6B

1. Quickest start is ollama run LiquidAI/lfm2.5-2.6b, 1.7 GB, or llama-cli -hf LiquidAI/LFM2.5-2.6B-GGUF
2. With vLLM or SGLang pass --tool-call-parser lfm2; for vLLM add --reasoning-parser qwen3 so the thinking comes back separately
3. Sample at temperature 0.1, top_k 50 and repetition_penalty 1.1, as the card and generation_config.json set
4. Put tools in the system prompt as List of tools: [...] JSON, or pass tools to apply_chat_template; calls come back as a Python list of function calls, and the system prompt can ask for JSON calls instead
5. Leave max_tokens unset or generous, since a low cap cuts the reasoning off mid-thought
6. Earlier thinking is dropped from the prompt unless preserve_thinking is true
7. Liquid's 328M LFM2.5-2.6B-DSpark drafter speeds decoding about 2.6 times in SGLang and on Apple silicon, per Liquid
8. Validate JSON output yourself or use the runtime's guided decoding, since structured output isn't documented