Laguna M.1 vs Qwen3-4B-Instruct-2507

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.

Qwen3-4B-Instruct-2507

A small, fast cousin for instruction following, short tool-calling jobs and long documents on modest hardware, with no thinking step. It runs in about 3 GB at 4 bits on every common runtime, and two providers host it. The main risk is neglect, with no file changes in over a year, no answers from Alibaba and no safety results.

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

Qwen3-4B-Instruct-2507

1. Quickest start is ollama run qwen3:4b-instruct, 2.5 GB. Plain qwen3:4b is the Thinking-2507 edition, not this one
2. On our reference machine use BF16 with vllm serve Qwen/Qwen3-4B-Instruct-2507 --max-model-len 262144 --enable-auto-tool-choice --tool-call-parser hermes
3. Use the hermes parser in vLLM or qwen25 in SGLang, the template emits JSON inside tool_call tags
4. Don't pass enable_thinking, this edition never thinks and the card says the flag isn't needed
5. Sampling per the card is temperature 0.7, top_p 0.8, top_k 20 and min_p 0, with presence_penalty up to 2 against repetition
6. Drop the context to 32,768 tokens if memory runs out, as the card suggests
7. Validate JSON yourself or use the runtime's guided decoding, since structured output isn't documented