Agents-A1

An open-weight (Apache-2.0) 35B mixture-of-experts agentic model from InternScience, released late June 2026. Its pitch — “Reaching Trillion-Parameter Performance with a 35B Agent” — is the spoke’s sharpest example yet of a lab arguing you close the frontier gap by scaling the agent horizon, not the parameter count. Self-reported numbers; treat as a first-party claim until an independent re-run (artificial-analysis) lands. arXiv tech report: 2606.30616; weights on Hugging Face.

What it is

35B MoE, total parameters; active-param count not disclosed in the repo/card.
262K context, served via SGLang or vLLM (the card ships a qwen3 reasoning parser, so it is built on or close to the Qwen3 architecture).
Agentic by design — built for long-horizon trajectories (the report cites ~45K-token average trajectories), function calling, and task decomposition across six heterogeneous domains unified into one model: long-horizon search, engineering, scientific research, instruction following, and tool-calling (the sixth being the general/agentic-reasoning base).

The method: multi-teacher domain-routed distillation

The claim is an efficiency argument, not a scale one. Rather than reinforcement learning, A1 uses a three-stage supervised paradigm:

Full-domain SFT — broad agentic behaviors across all domains.
Domain-level teacher models — a specialized expert trained per domain.
Multi-teacher, multi-domain on-policy distillation — the teachers are distilled back into one student with “heterogeneity-aware optimization” and salient-vocabulary alignment, so a single 35B model inherits six specialists’ competence.

So the lever is knowledge transfer + long agentic context, positioned against raising parameter count — a different mechanism from the spoke’s two existing efficiency levers (sparse MoE and quantization).

Self-reported benchmarks (vendor, snapshot)

Claimed parity with much larger frontier models (it names GPT-4.5 and DeepSeek-V4-pro as the “trillion-parameter” comparison):

GAIA (general agentic): 96.04
IFEval (instruction following): 94.82
Seal-0: 56.36
SciCode: 44.33
FrontierScience-Olympiad: 79.0 · FrontierScience-Research: 40.0

These are author-run; SciCode and GAIA also appear on the independent artificial-analysis yardstick, which is where the parity claim should eventually be checked.

Why it matters here

A1 sits with the open-weight challengers (qwen, deepseek, Z.ai) but adds a new wrinkle to the “does the frontier premium survive?” question: the squeeze may come not only from cheaper tokens but from training technique that buys frontier agentic behavior at mid-size. If it holds up independently, it strengthens the agentic-tooling thesis that capable open agents narrow the gap to closed frontier labs. Built on Qwen lineage, it’s also more evidence of Qwen as the open base layer others build on.

Cross-spoke

The agentic capability angle (long-horizon tool use, GAIA) is adjacent to agentic-tooling-wiki — but A1 is a model (weights + training method), so it’s the model market’s subject, not a harness.

internscience · moe-architecture · open-weight-models · qwen · artificial-analysis · deepseek · glm-52 · synthesis