Agents-A1-4B — a genuinely tiny agent model, shipped with its own quants

While Moonshot was preparing a 2.8-trillion-parameter flagship, InternScience went the other way: Agents-A1-4B landed July 14 — a 4.5B dense Apache-2.0 model post-trained specifically for agentic work, with 262K context and official Q4/Q8/F16 GGUFs published by the lab itself. It needs about 3 GB at Q4, which means it runs on a Mac mini, an old 8 GB card, or a laptop with no dedicated GPU at all. InternScience reports it beating its own Qwen3.5-4B base by wide margins on agent benchmarks. No planner pick yet — it is four days old — but this is the most interesting small model of the window.

Verdict: A 4B Apache-2.0 agent specialist with official GGUFs — the rare drop that runs on the laptop you already own

The take

The facts, verified against the Hugging Face repos via the HF API (`InternScience/Agents-A1-4B`, license tag apache-2.0, 4.54B params, created July 13 and announced July 14; the 35B-A3B flagship `InternScience/Agents-A1` landed June 26 with 33K downloads): the family is built for long-horizon search, engineering, scientific research, instruction following, and tool calling. Vendor numbers for the 4B: BrowseComp 66.8 (its Qwen3.5-4B base scores 47.2), XBench-DS 90.0, GAIA 95.1, IFEval 94.8, MatTools 49.3 (base: 10.9). Two things make those claims more credible than the usual launch-day table — InternScience open-sourced the evaluation framework so the protocol is reproducible, and they publish where the 4B loses (τ²-Bench 78.2 vs the base's 79.9), which vendors rarely do.

Read the framing honestly, though. The paper title — "Reaching Trillion-Parameter Performance with a 35B Agent" — is marketing, and the architecture strings confirm these are post-trains of Qwen3.5 bases (`Qwen3_5ForConditionalGeneration`), not from-scratch models. What they actually demonstrate is narrower and still interesting: heavy agent-specific post-training buys enormous gains on agent tasks at a fixed parameter count. A 4B that plans, calls tools, and searches competently is a different proposition from a 4B that chats.

Our call: model entry for the family, no planner pick yet. Our agents.low slot is Gemma 3 4B — proven, general, well-supported — and picks here follow real adoption, not launch-day tables (the 4B has a few hundred downloads so far). But this is the natural candidate to displace it: same size, Apache 2.0, purpose-built for exactly that slot, and the lab shipped its own GGUFs so llama.cpp/Ollama/LM Studio work day one. If independent numbers and community traction land by the next sweep, it moves. Meanwhile it is a genuinely low-risk thing to try — 3 GB, an afternoon, your own agent loop. The wider point: the same week the open frontier hit 2.8T parameters, the more useful release for most people was 4B.

Where this fits

Models: Agents-A1 (35B-A3B + 4B) · Gemma 3 (4B / 12B / 27B) · Qwen 3.5 4B · Qwen 3.6-35B-A3B

Hardware: Mac Mini M4 16 GB · NVIDIA RTX 3060 12 GB · RTX 5060 Ti 16 GB

Sources

Next step

Try this in the planner→