Dots3-Note Preview scores 76.8% on ARC-AGI-2

Dots Studio’s Dots3-Note Preview scored 76.8% on ARC-AGI-2, placing it first among open-weight models on the verified leaderboard when the result was published. ARC Prize estimated a cost of $0.08 per task.

ARC Prize ran the evaluation on a dedicated Baseten deployment with eight NVIDIA H100 GPUs and 640 GiB of total GPU memory. The reported cost uses Dots Studio’s API prices of $0.14 per million input tokens and $0.28 per million output tokens. It excludes Baseten hosting fees and therefore represents an API-price estimate rather than the cost of operating the evaluation hardware.

Dots Studio made the model available under the Apache 2.0 license on Hugging Face. The Dots API offered free access at the time of release.

A 280B model with a narrow active path

Dots3-Note Preview is a sparse mixture-of-experts model with 280 billion total parameters and 16 billion active parameters per token. A router selects eight of 256 specialized experts, alongside one shared expert, instead of running every parameter for each token. This design lowers inference computation while retaining a large pool of specialized weights.

The active parameter count describes how much of the network processes each token. Serving still requires enough memory to hold the full 280-billion-parameter checkpoint, which explains the recommended eight-GPU configuration.

Dots3-Note Preview also includes a 1.13-billion-parameter Multi-Token Prediction component. A compatible serving stack can use that layer as a draft generator for speculative decoding, proposing several tokens before the main model verifies them. The technique can reduce generation latency without changing the final model output.

The model is the smallest planned member of the dots3 family. Dots Studio has also announced jazz and aria variants aimed at different capability and cost tiers.

TEMPO extends the training horizon

Dots Studio introduced TEMPO, a reinforcement-learning method designed for agent tasks that can run for tens of hours. The same model alternates between an actor role, which advances the task, and a critic role, which checks progress and identifies unproductive behavior.

Periodic self-evaluation gives the training process intermediate signals before a long task ends. Dots Studio also released VibeSearchBench and VibeLifeBench, two environments intended to evaluate sustained, multi-step agent work.

One model, several input modes

The model card describes a broad set of supported workloads:

• Reasoning, coding and multi-step agent workflows
• Text, image, audio and video inputs with text-only output
• Tool calling through tool_choice
• Structured output through a JSON schema in response_format
• Long-document analysis within a 512K-token context window

A 512K context limit allows large document collections or codebases to fit in one request, although maximum capacity does not guarantee reliable recall across the full window. Developers should test retrieval accuracy, latency and memory use at the context lengths their applications require.

ARC-AGI-2 tests unfamiliar grid rules

ARC-AGI-2 consists of abstract visual grid problems that require a system to infer a transformation from a small set of examples and apply it to a new grid. The benchmark uses unfamiliar tasks to reduce the value of memorized answers from pretraining data.

The 76.8% result provides evidence of strong performance on that specific form of visual reasoning. It does not establish equivalent performance on coding agents, tool use or hours-long workflows, so those capabilities require separate evaluation. The reported eight-cent cost also depends on Dots Studio’s token pricing and should not be treated as a self-hosting estimate.

The reference deployment uses eight H100s

At release, vLLM supported the model on its main branch, while integrations for Transformers and SGLang remained under review. Dots Studio provided a development image for SGLang users in the interim. The recommended deployment uses the FP8 checkpoint on one eight-GPU node with SGLang or vLLM, matching the hardware class used for the verified ARC Prize run.

Basic text chat works through an OpenAI-compatible endpoint, allowing applications that use the OpenAI Python client to switch the base URL and model identifier:

from openai import OpenAI

client = OpenAI(
    base_url="http://127.0.0.1:8000/v1",
    api_key="EMPTY",
)

response = client.chat.completions.create(
    model="dots3-note-prev",
    messages=[
        {"role": "user", "content": "Hello!"}
    ],
)

OpenAI-compatible transport simplifies initial integration, though production testing should still cover tool schemas, structured-output compliance, multimodal payloads, timeouts and model-specific error handling.

What to measure before deployment

Teams considering Dots3-Note Preview can compare it with their current model across four practical areas:

• Task completion: Measure full workflow success, including recovery from failed tool calls.
• Long-context reliability: Test retrieval and reasoning at realistic document sizes instead of relying on the advertised limit.
• Serving economics: Include GPU utilization, batching, prompt length, output length and idle capacity.
• Operational control: Evaluate whether local weights, an Apache 2.0 license and private deployment justify the hardware requirements.

Dots3-Note Preview combines a verified ARC-AGI-2 score, downloadable weights, multimodal inputs and tooling for long-running agents. Its practical advantage will depend on whether those capabilities survive end-to-end testing on real workloads and whether the required infrastructure compares favorably with hosted alternatives.