Agents-A1: A 35B MoE Agentic Foundation Model

TL;DR

Agents-A1 is an open 35B-total MoE language model from InternScience built as a general agentic foundation model: long-horizon search, software-engineering, scientific research, instruction following, and tool calling are all targeted within a single base model. The project-page release ships SGLang and vLLM serving recipes with a 262K context window and Qwen3-style reasoning + tool-call parsers (suggesting a Qwen3-MoE base lineage). Training centers on a “domain-grounded knowledge-action graph” that converts agent process traces into trainable targets across heterogeneous domains, then transfers expertise between them. Sits in the same emerging niche as Qwen-AgentWorld-35B-A3B, Kimi K2.6, GLM-4.5, and Ornith-1.0-35B-MoE — open-weight ~35B MoE bases post-trained specifically for multi-turn tool-using agents rather than chat or coding alone.

Key claims

Single 35B-total MoE model is positioned as a heterogeneous agent foundation model covering long-horizon search, engineering, scientific research, instruction following, and tool calling rather than a single agentic domain [project page §Overview].
Training uses a “domain-grounded knowledge-action graph” that converts agent process traces into trainable targets, with three stated stages: alignment to broad agentic behaviors, per-domain expert capture, and cross-domain knowledge transfer [project page §Training].
Deployment ships as Qwen3-compatible — SGLang launch flags require --reasoning-parser qwen3 and --tool-call-parser qwen3_coder, and the vLLM recipe mirrors this — strongly implying a Qwen3-MoE base [project page §Serving].
Default context window is 262,144 tokens; both serving recipes are configured for tensor-parallel size 1, suggesting the active-parameter budget fits within a single high-memory GPU [project page §Serving].
The release bundles a model card, code repository, and “open evaluation framework for reproducible agent capability testing” — i.e. eval harness is shipped alongside the weights rather than referenced externally [project page §Resources].

Method

The project page is light on architectural detail; what is described is the post-training pipeline. Agent process traces from each target domain are structured into a domain-grounded knowledge-action graph: a representation where nodes capture domain state and edges capture executable actions, and the graph is sampled to produce supervision over multi-step decision sequences. Training proceeds in three stated phases — (1) align the base model with broad agentic behaviors (search, engineering, research, tools, instructions); (2) specialize per-domain experts to capture domain-specific reasoning; (3) transfer expertise across domains with optimization explicitly designed for heterogeneous knowledge transfer. The Qwen3-coder tool-call parser in the serving recipe implies tool-calling is emitted in the Qwen3-Coder XML-tag format rather than a custom schema.

Results

The project page does not publish benchmark numbers. The release ships an “open evaluation framework for reproducible agent capability testing” but does not surface specific scores on its landing copy. Headline framing — “scaling the horizon, not the parameters” — emphasizes long-horizon capability over raw parameter count, which is consistent with the ~35B-MoE size class but not directly measurable from what is published here.

Why it’s interesting

Lands in a tight cluster of open-weight ~35B-MoE agent foundation models that appeared in mid-2026 — OpenThoughts-Agent: Data Recipes for Agentic Models (OpenThoughts-Agent’s data recipe for the same niche), Qwen-AgentWorld-35B-A3B (Qwen-AgentWorld: Language World Models for General Agents, same 35B-A3B size class, also Qwen3-MoE-derived, training objective is environment simulation rather than action), and the ROME/ALE agent stack (Let It Flow: Agentic Crafting on Rock and Roll, Building the ROME Model within an Open Agentic Learning Ecosystem, Qwen3-MoE base + full agentic-RL ecosystem). Agents-A1’s contribution within this cluster is the knowledge-action graph supervision framing — process traces as graphs rather than as flat trajectories — and the explicit cross-domain transfer stage, which is a different bet than Qwen-AgentWorld’s next-state-prediction objective or Ornith-1.0’s self-scaffolding RL. Worth filing because the design space for “what does the right supervision look like for an open 35B agentic base” is actively contested right now and each new datapoint constrains it.