Agents-A1-4B vs Qwen3-4B-Instruct-2507

Agents-A1-4B

A small agent cousin for search, research and tool-calling loops, with vision and 262,144 tokens of context. It runs in about 4 GB from InternScience's Q4_K_M GGUF, in vLLM, SGLang, Transformers or llama.cpp. The main risks are an undeclared base, no safety results or agent guidance, and a template that adds its own dated system prompt when you send none.

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.

Agents-A1-4B

1. Quickest start is llama-server -hf InternScience/Agents-A1-4B-Q4_K_M-GGUF --jinja, about 3.4 GB with the vision projector
2. With vLLM use --reasoning-parser qwen3 --enable-auto-tool-choice --tool-call-parser qwen3_coder, as on the card, and add --language-model-only for text-only work
3. Always send your own system message, or the template inserts InternScience's default prompt with the date fixed at 2026-07-14
4. The card recommends temperature 0.85, top_p 0.95, top_k 20, min_p 0 and presence_penalty 1.1, and there is no generation_config.json to set them
5. Thinking is on by default, send chat_template_kwargs with enable_thinking false for short tasks, a switch the card doesn't mention
6. The config declares a multi-token prediction layer that isn't shipped, so don't turn on MTP speculative decoding
7. Validate JSON output yourself, 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