Shanghai AI Lab models: Intern-S2 to Sage-A1.5 compared

Short answer: Intern-S2 for images, Atria-Dawn for long text, Agents-A1 and Sage-A1.5 for agents

LLMTR is a gateway that gives access to language models from different providers through one OpenAI-compatible API. Shanghai AI Lab's five models, Intern-S2 (internlm/intern-s2-free), Atria-Dawn (internlm/atria-dawn-free), Atria-Dawn Preview (internlm/atria-dawn-preview-free), Agents-A1 (internlm/agents-a1-free) and Sage-A1.5 (internlm/sage-a1.5-free), are called on the /v1/chat/completions endpoint at https://llmtr.com/v1 with the same API key from your llmtr.com account, and they are free until 31 October 2026, 23:59 (TRT). Use Intern-S2 to read a scientific figure or table from an image, Atria-Dawn for long text-based agent work that needs a 524,288-token context, Agents-A1 for tool-using agents that also take images, and Sage-A1.5 for flows that split work across subagents.

Atria-Dawn Preview is the MIT-licensed open preview of Atria-Dawn. All five support tool calling and reason before answering by default. From 1 November 2026, 00:00 (TRT), the five ids return 410 model_retired and are not redirected to another model.

Which of Shanghai AI Lab's five models fits which job?

The table shows the LLMTR catalog's card values.

With tools attached, tool_choice none returns 400 on Intern-S2 and both Atria rows; when a request uses no tool, leave the tools field out.

What is each model for?

According to Shanghai AI Lab's model card, Intern-S2 is a multimodal foundation model for scientific intelligence and long-horizon agents; through visual pretraining it learns from raw pages of scientific literature, and it was trained on scientific reinforcement-learning tasks across more than 20 domains. It suits a question sent with an image of a chart, figure or table; ask for JSON in the prompt and use auto for tool_choice.

According to its card in the LLMTR catalog, Atria-Dawn is an agent model that completes multi-step research and engineering tasks with tools, running analysis, solution design, coding, experiment runs, result review and recovery from failures in one loop. Its context window is 524,288 tokens; when you send a long document set in one request, the prompt and the answer combined cannot exceed what is left of the day's allowance.

Atria-Dawn Preview is the preview of the same model, with weights open under the MIT license; it lets you see the output of the open release you may run on your own servers through the same API first.

According to InternScience's model card, Agents-A1 is an Apache 2.0, 35B mixture-of-experts (MoE) agent model scaled across long-horizon search, engineering, scientific research, instruction following and tool calling. Per the LLMTR catalog, 3 billion of its parameters are active per token and it accepts images.

Sage-A1.5's model card presents it as a 35B MoE agent model built on Qwen3.6 and Agents-A1 that learns to organize work and revise its approach during execution, running subagents with isolated contexts in parallel. According to the card, this mode is turned on when you define a tool named spawn_subagent; the weights are under Apache 2.0.

• Shanghai AI Lab, Intern-S2 model card · 2026-10-22
• InternScience (Shanghai AI Lab), Agents-A1 model card · 2026-10-22
• InternScience (Shanghai AI Lab), Sage-A1.5 model card · 2026-10-22

Trying all five with the same prompt

The Python example below sends the same short task to the five models with reasoning off and prints finish_reason and usage.total_tokens for each. With no image and no tools, the request is valid on all five rows. If you leave reasoning on, raise max_tokens above 4,096: reasoning tokens count toward it, and a short limit can cut the answer off with finish_reason length.

OpenAI Python SDK example that sends the same prompt to the five Shanghai AI Lab models through https://llmtr.com/v1/chat/completions with reasoning off and prints finish_reason and the total token count.

import os

from openai import OpenAI

client = OpenAI(base_url="https://llmtr.com/v1", api_key=os.environ["LLMTR_API_KEY"])

# The five Shanghai AI Lab rows: daily token allowance until 31 October 2026,
# 23:59 (TRT); from 1 November 2026 they answer 410 model_retired.
MODELS = [
    "internlm/intern-s2-free",
    "internlm/atria-dawn-free",
    "internlm/atria-dawn-preview-free",
    "internlm/agents-a1-free",
    "internlm/sage-a1.5-free",
]
messages = [{"role": "user", "content": "List three checks to run before merging a database migration, one line each."}]

for model in MODELS:
    response = client.chat.completions.create(
        model=model,
        messages=messages,
        max_tokens=4096,
        # Reasoning off; reasoning tokens would count toward max_tokens and the allowance.
        extra_body={"reasoning": False},
    )
    print(model, response.choices[0].finish_reason, response.usage.total_tokens)
    print(response.choices[0].message.content)

Rules shared by all five: free period, daily allowance, reasoning and data

The five models are free until 31 October 2026, 23:59 (TRT), and nothing is charged to your balance. The allowance is separate for each account and each model: 500,000 tokens a day, counting input and output combined, reasoning tokens included, and new accounts start with a smaller allowance on their first day; it renews daily at 00:00 (TRT). When it runs out, the request returns 429 rate_limit_error and the Retry-After header tells you when to try again; if the prompt is larger than what is left, the error tells you to send a shorter request.

All five reason before answering, and the reasoning text arrives in the reasoning_content field. To turn it off, send "reasoning": false in the request (extra_body={"reasoning": False} in Python) or add :fast to the model id, for example internlm/agents-a1-free:fast; :think or "reasoning": true keeps it on. These rows have no levels, only on and off, so reasoning_effort returns 400.

Streaming is supported; send "include_usage": true in stream_options and usage arrives in the last chunk. There is no cache discount, and output is bounded by the context window. The five models are also listed in the model picker of LLMTR Chat (chat.llmtr.com).

Data notice: these rows are processed through a third-party provider. Prompts and completions may be logged by the provider and used to improve the model; do not send sensitive, personal or customer data.

Which row to use from 1 November 2026

The five models are free until 31 October 2026, 23:59 (TRT); from 1 November 2026, 00:00 (TRT), requests to the five ids return 410 model_retired and are not redirected to another model. There is no successor; change the model id in your code yourself.

After 31 October 2026, Intern-S2's scientific questions with images move to the free Nemotron 3.5 Super VL or the paid Qwen3.6 35B-A3B or DeepSeek V4.1 Flash. After 31 October 2026, Atria-Dawn and Atria-Dawn Preview move to the free Nemotron 3 Ultra or Agnes 2.5 Flash, or the paid DeepSeek V4.1 Flash or MiniMax M3.1 Flash. After 31 October 2026, Agents-A1 and Sage-A1.5 move to the paid Qwen3.6 35B-A3B, in the same size class, or the free Nemotron 3.5 Super VL.

response_format, tool_choice and reasoning settings also differ by row, and Agnes 2.5 Flash takes images only as base64 data URLs. The series article on the period after 1 November has these rows' model ids and the details.

Which one when?

The choices follow from the differences in the table; compare output quality with your own examples. Do not use any of these five models for work with sensitive, personal or customer data: their data notices say prompts and completions may be logged by the provider and used to improve the model.

• Reading a scientific figure, chart or table from an image when you can ask for JSON in the prompt: Intern-S2 (internlm/intern-s2-free); it accepts images and is free until 31 October 2026, 23:59 (TRT).
• A text-based research or engineering agent that needs a 524,288-token context, structured output with json_schema and required tool_choice: Atria-Dawn (internlm/atria-dawn-free).
• Trying the MIT-licensed open release you may run on your own servers through the same API first: Atria-Dawn Preview (internlm/atria-dawn-preview-free); for the stable release, Atria-Dawn.
• A long-horizon tool agent that also reads images and works with auto or a named tool_choice: Agents-A1 (internlm/agents-a1-free).
• A flow that hands work to subagents with isolated contexts: Sage-A1.5 (internlm/sage-a1.5-free); according to its model card, this mode is turned on when you define a tool named spawn_subagent.

Comparing the five Shanghai AI Lab models on the same task

Send the same task to the five models with one LLMTR key, look at finish_reason and the token count, and choose the id that fits your work.

1. Key. Open an account on llmtr.com and create a key under Dashboard > API Keys.
2. Same task. Send your own task with the five model ids to the https://llmtr.com/v1/chat/completions endpoint with reasoning off.
3. Choice. Compare the answers, finish_reason, the token count and the data notice, choose the id, and pick the row that replaces it before 1 November 2026.

Frequently asked questions

Until when are the Shanghai AI Lab models free on LLMTR?

Until 31 October 2026, 23:59 (TRT). Each account gets 500,000 tokens a day per model, input and output combined; new accounts start with a smaller allowance on their first day. From 1 November 2026, 00:00 (TRT), requests return 410 model_retired.

How do you turn reasoning off on these models?

Send "reasoning": false in the request or add :fast to the model id. reasoning_effort returns 400 on these rows. Reasoning tokens count toward max_tokens and toward the daily allowance.

Which models accept images?

Intern-S2, Agents-A1 and Sage-A1.5 accept images alongside text (URL or base64). Atria-Dawn and Atria-Dawn Preview take text only; a request with an image returns 400.

Related posts

• Reading scientific charts and tables with the Intern-S2 API
• Atria-Dawn vs Atria-Dawn Preview: stable release or open preview
• Agents-A1 vs Sage-A1.5: which agent model for which job
• Building a tool-calling agent loop with Agents-A1
• Reasoning and the daily token allowance on Shanghai AI Lab models
• Long-document extraction with Atria-Dawn and JSON Schema
• Intern-S2 vs Qwen3.8 27B, Nemotron 3.5 Super VL, Agnes 2.5 Flash
• Agents-A1, Sage-A1.5 and Qwen3.6 35B-A3B compared
• Shanghai AI Lab models after 1 November 2026: migration guide

• Intern-S2 model
• Atria-Dawn model
• Atria-Dawn Preview model
• Agents-A1 model
• Sage-A1.5 model
• Reasoning effort documentation
• Tool calling documentation
• Multimodal images
• Errors documentation