DiffusionGemma 26B A4B IT vs Qwen3-Coder-Next

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-Coder-Next

A fast coding cousin for agent harnesses, with 3B active parameters and tool calling trained across many scaffold formats for Claude Code, Cline or Qwen Code loops. Alibaba's 4-bit GGUF needs about 58 GB, fitting one 64 GB machine, and two providers host it. The main risk is neglect, with no change since February, no replies and no safety results.

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-Coder-Next

1. Quickest start is ollama run qwen3-coder-next (Q4_K_M, 52 GB of files)
2. On our reference machine Alibaba's Q8_0 GGUF (84.8 GB, about 102 GB in use) also fits, run with llama-server -m Qwen3-Coder-Next-Q8_0/Qwen3-Coder-Next-Q8_0-00001-of-00004.gguf --jinja -ngl 99 -fa on -c 131072
3. Serve with vllm serve Qwen/Qwen3-Coder-Next --enable-auto-tool-choice --tool-call-parser qwen3_coder (vLLM 0.15.0 or later), using the built-in parser rather than the repository's qwen3coder_tool_parser_vllm.py, which fails to import on newer vLLM per discussion 36
4. The card's commands use tensor parallel 2 though the text says 4 GPUs, and it suggests cutting the context to 32,768 if the server fails to start
5. There's no thinking mode, so don't pass enable_thinking or wait for reasoning_content
6. Sampling per the card is temperature 1.0, top_p 0.95 and top_k 40
7. The GGUF card's llama.cpp example path omits the quant from the file name; the real files are named like Qwen3-Coder-Next-Q4_K_M-00001-of-00004.gguf
8. Validate any JSON it returns, since structured output isn't documented