Gemma 4 E2B IT vs Laguna M.1

Gemma 4 E2B IT

A small cousin for light agent work, tool-calling loops, transcription and image reading on a phone or laptop, under Apache 2.0. Google's 4-bit build needs about 5 GB with working space and runs in all the common runtimes. Google's card puts it far below larger Gemma 4 sizes on agent and coding tasks, with no published safety results.

Laguna M.1

A large coding and agent cousin under plain Apache 2.0, for agents with a multi-GPU server. Its smallest build needs about 167 GB, more than a 128 GB machine, and no local runtime library packages it. The main risks are no upkeep since July, no hosted fallback shown today, no safety results, and repository code on the SGLang route.

Gemma 4 E2B IT

1. On the reference machine run llama-server -hf google/gemma-4-E2B-it-qat-q4_0-gguf --hf-file gemma-4-E2B_q4_0-it.gguf --jinja -c 131072, and add --mmproj with gemma-4-E2B-it-mmproj.gguf for images and audio
2. With Ollama run ollama run gemma4:e2b and raise num_ctx or OLLAMA_CONTEXT_LENGTH, since the default context is small and agents want at least 64K
3. In vLLM or SGLang pass --tool-call-parser gemma4 --reasoning-parser gemma4 (vLLM also --enable-auto-tool-choice), or tool calls come back as text; Google's w4a16 build is the 4-bit route there
4. Thinking is off by default; pass enable_thinking=True or start the system prompt with <|think|>, and drop earlier turns' thinking from history except on tool-call turns
5. Put images before the text and audio after it, keep audio to 30 seconds and video to 60 frames, and validate tool names and arguments before running them

Laguna M.1

1. Run the NVFP4 build on Blackwell GPUs with vllm serve poolside/Laguna-M.1-NVFP4 --tensor-parallel-size 2 --tool-call-parser poolside_v1 --reasoning-parser poolside_v1 --enable-auto-tool-choice, on vLLM 0.21.0 or later
2. Thinking is off in the template unless you pass enable_thinking true or start the server with --default-chat-template-kwargs '{"enable_thinking": true}', which Poolside recommends for agentic coding
3. Pass earlier reasoning back as reasoning_content in the history, since the card says reasoning may stop in later steps if it's dropped
4. Sampling per the card is temperature 1.0 and top_k 20
5. The FP8 build's Docker image and serve.sh predate upstream support and set a 4,096-token context, so prefer a released vLLM
6. It doesn't fit our 128 GB reference machine at any published precision
7. Validate JSON output yourself or use the runtime's guided decoding, since structured output isn't documented