DiffusionGemma 26B A4B IT vs Qwen3.5-397B-A17B

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.

Qwen3.5-397B-A17B

Alibaba's biggest open Qwen3.5, a cousin for hard coding, agent and vision work. Even at 4 bits it needs about 270 GB, so most agents will call it through Alibaba's Model Studio or one of the many providers that host it. The main risk is silence, with no safety results, no agent guidance and template reports without a reply.

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

Qwen3.5-397B-A17B

1. The card's server command is vllm serve Qwen/Qwen3.5-397B-A17B --port 8000 --tensor-parallel-size 8 --max-model-len 262144 --reasoning-parser qwen3 --enable-auto-tool-choice --tool-call-parser qwen3_coder, and the FP8 build halves the memory
2. It doesn't fit our 128 GB reference machine at any precision, so call a hosted endpoint instead (Model Studio's Qwen3.5-Plus, or a provider serving Qwen/Qwen3.5-397B-A17B)
3. On Model Studio pass "enable_thinking": false directly, while vLLM and SGLang take "chat_template_kwargs": {"enable_thinking": false}
4. Sampling per the card is temperature 0.6, top_p 0.95 and top_k 20 when thinking, and temperature 0.7, top_p 0.8, top_k 20 and presence_penalty 1.5 without
5. For multi-token prediction in vLLM add --speculative-config '{"method":"qwen3_next_mtp","num_speculative_tokens":2}'
6. Keep the context at 128K or more if you rely on thinking, and turn on YaRN only past 262,144 tokens
7. Validate any JSON it returns, since structured output isn't documented