Agents-A1
35B of agent, distilled from a room full of teachers
What it is
Shanghai AI Lab's 35B-total MoE agentic model (with a 4B sibling), trained via multi-teacher, multi-domain on-policy distillation across long-horizon search, engineering, and scientific research. 262K context, native tool calling, vision-language input, and llama.cpp/Ollama builds for local use.
Why it's interesting
The strongest small open agent model of the summer — state-of-the-art on SEAL-0 (56.4) and GAIA (96.0), claiming parity with trillion-parameter systems from a model that fits on a single node, under plain Apache-2.0 with no rider.
Use cases
- Self-hosted deep-research and search agents
- Tool-calling backends at modest hardware cost
- Distillation and agent-training research
Who it's for
Agent builders without frontier budgets; researchers
Setup
Easy. Single node via vLLM or SGLang; GGUF quants for llama.cpp, Ollama, and LM Studio; the 4B variant reaches laptops
Limitations & cautions
The parity-with-flagships claims are self-reported and independent evaluation is still sparse; gains concentrate in search and science domains rather than general chat.
Editorial takeaway
The agent race's counter-thesis in one release: maybe you don't need a trillion parameters, just better teachers.