SSyncDev

Sage S2.1

SyncDev model directory · Profile updated Nov 18, 2026

Compare with another model

Sage S2.1 (released as Atria Dawn) is Shanghai AI Lab's open 744B MoE agent model on GLM-5.2, with a 512K context and a hosted API.

Key specifications

Specification score
53.3
Context window
512K
License
MIT

The specification score summarizes recorded features and limits. It is not a benchmark result.

Pricing conditions: No per-token price published with the weights; hosted access is via the Intern Discovery Token Plan. Confirm current terms in your console.

Inputs and outputs

Recorded modality support shows what this model can accept and produce. An undocumented modality is shown as unknown rather than assumed to be unsupported.

Accepts

  • Text input

Produces

  • Text output

Capabilities

ReasoningFunction callingOpen weights

What is Sage S2.1?

The naming is the first thing to sort out. Shanghai AI Laboratory shipped this model as Atria Dawn, and the Atria Dawn repository now redirects to internlm/Sage-S2.1. The weights page still says Atria Dawn throughout and tells you to send Atria-Dawn as the model value in API calls. Use that string until the API console tells you otherwise. Sage S2.1 and Atria Dawn are the same release.

It is an agentic model built on the 744B-parameter mixture-of-experts GLM-5.2. The aim is narrow: push open-ended research and engineering work to results that can be executed, verified and reproduced, including recovery when a step fails.

Four kinds of work

  • Discovery: deep research, evidence gathering, turning a question into an experiment plan.
  • Creation: software, interactive apps, games, visualisations, machine-learning systems.
  • Delivery: reports, presentations and other structured documents from raw data and requirements.
  • Cybersecurity: analysing, validating and fixing vulnerabilities in authorised environments.

Key facts

Field Value
Hub repo internlm/Sage-S2.1 (FP8 build available)
Released 23 October 2026; Atria Dawn Preview on 11 September
Context 512K (preview: 256K)
API model value Atria-Dawn
Endpoints OpenAI-compatible and Anthropic-compatible
Licence MIT

Publisher-reported scores

Against DeepSeek V4 Pro 0813, Kimi K3, Qwen 3.8 Max, GLM 5.3 and others, the lab reports AutomationBench 54.7, CyberGym 86.8, SkillsBench 67.8 and WideSearch 84.6. It does not win everywhere: GLM 5.3 leads Workspace-Bench-Lite and Qwen 3.8 Max leads Workspace-Bench in the same table. That honesty is useful. It tells you this is an automation and security specialist, not a universal winner.

Hosted access and pricing

The lab offers hosted access through its Intern Discovery Token Plan, with international and China consoles, plus ready-made config snippets for Codex (Responses API) and Kimi Code. No per-token price is published with the weights, so check the console for current plan terms. Set the client's context limit to 512,000 so tools do not truncate early.

Before you adopt it

Self-hosting a 744B MoE is a cluster job, FP8 or not. Most teams should start on the hosted endpoint, run a week of their real automation tickets, and log success rate and retries. Compare with the faster Sage S2.1 Flash if latency matters more than the last few points.

How to evaluate Sage S2.1

Compare the model on the work you need to complete.

Tasks and inputs

Sage S2.1 is listed for Text generation, Coding & Code Generation, Computer Use. It accepts text and produces text. Use the publisher’s input format and preprocessing for your chosen deployment.

Capacity and cost

Documented limits: 512K context. Measure a completed task using the prompts, documents and tools your application will use. Include reasoning and retries in the result.

Choosing a deployment

The weights are downloadable under MIT. Check runtime support and size the complete model and working context for your expected concurrency.

Frequently asked questions

Yes. The Atria Dawn repository redirects to internlm/Sage-S2.1, and the weights page still describes the model as Atria Dawn.

These records share a directory category. Check the source card, output type and deployment requirements before treating a model as a replacement.

Microsoft

Aion 1.0 Instruct 2.3B

Aion 1.0 Instruct 2.3B is Microsoft's MIT-licensed on-device small language model and Phi Silica successor for Windows, with a 32K context.

32.77K context

Google

Gemini 4 Neon

Gemini 4 Neon is Google's mid-tier Gemini 4 model for complex coding and analysis, sitting between Hydrogen and Carbon at $3/$15 per 1M tokens.

1.05M context

OpenAI

GPT-6.2 Astra

GPT-6.2 Astra is OpenAI's November 2026 flagship for hard professional, scientific and agentic work, at $8/$40 per million tokens with a 1.05M context.

1.05M context

OpenAI

GPT-6.2 Sol

GPT-6.2 Sol is OpenAI's everyday model for coding, agents and writing: $2/$10 per million tokens, 1.05M context, released 17 November 2026.

1.05M context

Alibaba

Qwen4 Max

Qwen4 Max (qwen4-max) is Alibaba's first Qwen4-generation flagship: 1M-token context, text, image and video input, thinking on by default, $1.50/$4.50 per 1M.

1M context

Sarvam

Sarvam-105B-1711

Sarvam-105B-1711 is a refreshed open post-train of Sarvam's 105B MoE, tuned for tool calling, instruction following, agents and Indian languages.

— context