RepoRadar
Leave a Tip
Item detail
github.com

InternScience/Agents-A1

InternScience/Agents-A1 is a 35b moe agentic model targeting that RepoRadar is tracking in its New Models section, currently rated Gold tier with a 'try now' verdict. Its strongest signal is workflow potential, scored 9.2 out of 10.

Open source →Back to radar
Score8.1
Popularity1.0
Risklow
TierGold
Score breakdown
Usefulness9.0
Novelty8.0
Momentum8.0
Maturity6.4
Open-source/build8.4
Evidence7.2
Workflow potential9.2
Setup ease6.4

Popularity is tracked separately. Support, ads, sponsorships, and tips never affect these signals.

Why it matters

Useful for any developer, research team, or organization that wants to evaluate or deploy a 35B-A3B MoE agentic model with open weights — the combination of Apache-2.0 + open weights + 50-author paper + six-domain evaluation + Apple-Silicon-local-friendly .mlx ports + ModelScope mirror makes this a credible alternative to OpenAI / Anthropic / Google proprietary agentic models for teams with the ri

Who should use it

Any developer or research team that wants to evaluate or deploy an open-weight 35B-A3B MoE agentic model with Apache-2.0 license — the open weights + Apache-2.0 + 50-author paper + six-domain evaluation + Apple-Silicon-local-friendly .mlx ports + ModelScope mirror makes this a credible alternative to OpenAI / Anthropic / Google proprietary agentic modelsAnyone who values the horizon-scaling argument over the parameter-scaling argument — the paper's core claim is that scaling the agent horizon (long-horizon trajectories + heterogeneous agent abilities via a knowledge-action infrastructure) reaches trillion-parameter-level performance in a 35B-A3B MoE, a meaningful empirical argument against the 'just make it bigger' frontier-model playbookAnyone who needs the multi-teacher domain-routed on-policy distillation recipe — the training recipe is well-documented and reproducible to the degree that any team with 50+ GPUs can replicate itAnyone who needs six-domain evaluation breadth — long-horizon search (HLE, BrowseComp, GAIA), engineering tasks (SWE-Bench-style), scientific research (SciCode, FrontierScience-Olympiad, FrontierScience-Research, HLE, MolBench-bind), instruction following (IFEval, IFBench), general agentic tasks (HiPhO, Seal-0, XBench-DS-2510), scientific agentic tasks (FrontierScience-Research, MolBench-bind)Anyone who needs Apple-Silicon local inference — mlx-community .mlx ports let Apple Silicon users run Agents-A1 on a Mac without an external GPUAnyone who needs Chinese-language developer coverage — the ModelScope mirror serves the Chinese-language developer audience, which is a non-trivial share of the open-weight agentic-modeling communityAnyone who needs resource-constrained deployment — Hugging Face quantized variants (GPTQ / AWQ / BNB) cover the practical deployment spectrum: GPTQ/AWQ for NVIDIA, BNB for resource-constrained finetuningAnyone who needs a 35B-A3B MoE with 45K-token average trajectory training data — long enough for internalizing real agentic reasoning, short enough to fit on commodity multi-GPU

Who should skip it

Skip InternScience/Agents-A1 unless the captured evidence suggests it solves a problem you are actively working on.

About this signal

InternScience/Agents-A1 is tracked by RepoRadar as a 35b moe agentic model targeting in the New Models section. It was first seen on 2026-07-04 and last updated on 2026-07-04. The current verdict is 'try now' with a Gold tier and moderate setup difficulty. InternScience/Agents-A1 leads on workflow potential (9.2) and practical usefulness (9.0); its lowest signal is setup ease (6.4), so factor that in before investing setup time. This page summarizes the evidence RepoRadar has captured from captured source metadata. The score, tier, risk label, and verdict on this page are never influenced by sponsorship, ads, or tips — they reflect only the usefulness, popularity, novelty, momentum, maturity, and evidence signals described in the RepoRadar methodology.

How this item is evaluated

RepoRadar assigned InternScience/Agents-A1 a composite score of 8.1 out of 10, placing it in the Gold tier. This score combines weighted sub-signals: usefulness (35%), novelty (18%), momentum (14%), maturity (10%), open-source/build quality (7%), evidence quality (6%), workflow potential (6%), and setup ease (4%). Popularity is tracked separately at 1.0 and never affects the composite score or tier. The risk label of 'low' reflects inherent user-impacting hazards, not generic novelty. Items with no risk flag may still require normal code review before production use.

Putting this into practice? Read How to vet an AI agent or MCP server before you wire it in for the checklist behind this score.

Risk explanation

README's comparison table (the `Larger-scale Models` column in the per-benchmark tables) references model names that are not currently publicly released as of the cycle date (`GPT-5.5(xhigh)`, `DeepSeek-V4-pro(Max)`, `Kimi-K2.6`); the cycle 146 fictional-model-name-forward-looking rule treats this as a `risk_flag` + `conditional` verdict because the project is a real, runnable, open-weight 35B-A3B MoE agentic model that ships reproducible multi-domain benchmarks against the currently public Qwen3.5-35B-A3B + Step-3.5-Flash + comparable-model baseline; Apache-2.0 license with open weights + 50-author paper + Hugging Face collection + ModelScope mirror + mlx-community ports — the combination is the right open-source shape; verify the LICENSE before any commercial embedding that might trigger the cycle 126 'Modified Apache with commercial-use caveat' pattern (this repo is plain Apache 2.0, confirmed 2026-07-04).

Evidence links
Closest alternatives / related signals
open-weight-modelagentic-model35b-moe35b-a3bmoe-mixture-of-expertshorizon-scalinglong-horizon-trajectoriesheterogeneous-agent-abilities