Agents-A1: a 35B-A3B agentic model that fits a 24GB card

Agents-A1 is an agentic post-train of Qwen3.5-35B-A3B by InternScience, released 26 June 2026 under Apache 2.0 - no acceptable-use policy, no user cap, no pass-through duty. It is aimed at long-horizon search, engineering, scientific research, instruction following and tool calling, and it is small enough to run on one consumer GPU.

It is a Qwen3.5 post-train, and says so

config.json reports model_type: qwen3_5_moe and Qwen3_5MoeForConditionalGeneration, with 40 layers, 256 experts, 8 active per token and a 2048 hidden size. InternScience’s own comparison table lists Qwen3.5-35B-A3B and Qwen3.6-35B-A3B as the “comparable models” it is measured against.

That is not a criticism. It is the useful fact: if you already run Qwen3.5-35B-A3B locally, this is the same shape, the same VRAM budget, and the same tooling - with agent behaviour trained in. The training recipe is full-domain SFT, then domain-level teacher models, then multi-teacher on-policy distillation back into one model.

Where it lands

Against frontier models several times its size:

InternScience claims overall SOTA on Seal-0, HiPhO, both FrontierScience splits, IFBench and IFEval, and best-in-class at ~35B on the rest. Self-reported, as always - but the comparison set is GPT-5.5, DeepSeek-V4-pro and Kimi-K2.6, not a friendly field.

Settings

The model card is unusually specific, and these are not the greedy defaults you would guess for an agent:

temperature       0.85
top_p             0.95
top_k             20
presence_penalty  1.1

The presence penalty is the interesting one - it exists to stop long-horizon agent loops from re-issuing the same tool call forever.

What fits

22.1 GB for the 35B with vision is tight on a 24GB card once you want context, and comfortable on 32GB. The 4B variant landed 14 July and at 3.4 GB total it is the one to put on a laptop or a second card as a cheap tool-calling worker. FP8 and NVFP4 builds exist for vLLM, and mlx-community has Mac quants.