DiffusionGemma 26B A4B IT vs Hy3

DiffusionGemma 26B A4B IT

A fast local generation cousin for low-concurrency work on text, images and video, with optional thinking and Gemma 4's tool format, under Apache 2.0. It needs about 62 GB for Google's BF16 files and runs only in Transformers or SGLang. The main risk is quality, since Google's own figures put it below Gemma 4 26B A4B, and it's labelled experimental.

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.

DiffusionGemma 26B A4B IT

1. Serve it with sglang serve --model-path google/diffusiongemma-26B-A4B-it --dllm-algorithm Gemma4Renoise --trust-remote-code on SGLang v0.5.22 or later, then call /v1/chat/completions
2. In Transformers, install transformers>=5.11.0 and load DiffusionGemmaForBlockDiffusion with dtype="auto"; generation_config.json already holds the recommended sampler
3. Turn thinking on with enable_thinking=True in apply_chat_template or <|think|> at the start of the system prompt, read the reasoning between <|channel>thought and <channel|>, and drop it from history before the next turn
4. Don't rely on temperature, top_p, logit_bias or JSON Schema in SGLang, which ignores or rejects them for this model; validate JSON yourself
5. Streaming arrives one 256-token canvas at a time, not token by token
6. Put images before text in the prompt, and pick an image token budget of 70 to 1,120 by how much detail the task needs

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