DiffusionGemma 26B A4B IT vs Swift 1.5 Qwen3.8-27B

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.

Swift 1.5 Qwen3.8-27B

A Qwen3.8-27B cousin post-trained to think in fewer tokens, for coding and agent work with images in, that fits a 24 GB machine with UkisAI's own 4-bit builds. Commercial use is free only below US$1M revenue. The main risks are that nobody at UkisAI answers its discussions, and that the shorter thinking rests on UkisAI's own benchmarks.

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

Swift 1.5 Qwen3.8-27B

1. Quickest start is llama-server -hf ukisai/Swift-1.5-Qwen3.8-27B-GGUF:Q4_K_M --jinja, about 18.4 GB with the vision projector
2. On our reference machine BF16 fits, vllm serve ukisai/Swift-1.5-Qwen3.8-27b --max-model-len 262144 --reasoning-parser qwen3 --enable-auto-tool-choice --tool-call-parser qwen3_coder
3. The template is Qwen3.8-27B's, so chat_template_kwargs with reasoning_effort low, medium or xhigh (default) and enable_thinking false work as on the base
4. UkisAI benchmarked it at temperature 1.0, top_p 0.95, top_k 20 and presence_penalty 0, which is a reasonable starting point
5. The MLX 4-bit build needs UkisAI's patched MLX-LM and has no image or video chat
6. Check the licence before commercial use, it stops at US$1M gross annual revenue including affiliates
7. MTP self-speculative decoding is available with the base's head, flags on the card