Skip to main content

Baseten releases two decision models with 128K context and image input

21 NOVEMBER 2026·2 MIN READ·4 SOURCES

Baseten says it released bison-1 and groundhog-1 on November 20, 2026. Both models support context windows of up to 128,000 tokens and accept text and images as inputs.

Key takeaways · 3

  • 01

    Both models support up to 128,000 tokens of context, which Baseten says is eight times the capacity of its previous Model API offering.

  • 02

    Both models accept text and images, and use the same API as Baseten’s existing decision models.

  • 03

    Baseten positions bison-1 for accuracy and groundhog-1 for lower latency and cost.

Two models for predefined decisions

Baseten says it released bison-1 and groundhog-1 on November 20, 2026, extending its decision-model work with two models built on Google’s Gemma 4.[1][2] The company describes the models as returning calibrated probabilities for predefined decisions.[1] Baseten says the release brings its decision-model work to an American backbone.[2] That description and the stated model capabilities are the company’s framing; the supplied information does not specify which decision tasks the models support or provide results from customer deployments. For teams evaluating them, the practical next step is to test their own predefined decisions and inputs rather than assume a benchmark result applies to their workflow.

Longer context and image input

Both models support context windows of up to 128,000 tokens, and Baseten says that is eight times the capacity of its previous Model API offering.[2] They also accept both text and images as inputs.[2] These features may make the models worth evaluating for workflows that combine lengthy material with visual inputs, but the announcement does not provide examples of specific tasks or report how performance changes with longer inputs. Teams should test whether their own documents, images and decision definitions fit the models’ capabilities before choosing a deployment approach.

Different model profiles and benchmark result

Baseten describes bison-1 as a 31-billion-parameter dense model focused on accuracy.[2] It describes groundhog-1 as a 26-billion-parameter mixture-of-experts model with roughly 4 billion active parameters, designed for lower latency and cost.[2] Baseten also says it recorded the index’s lowest Brier score, a measure of how closely confidence tracks correctness.[2] Those benchmark claims do not establish outcomes on a particular organization’s decisions.

Where teams can access them

Baseten says both models are available on Hugging Face and through Baseten Model APIs, and that they use the same API as its existing decision models.[2] The company says bison-1 can be used instantly with per-token pricing.[3] Hugging Face lists bison-1’s license as MIT and Apache 2.0.[4] The supplied information does not state a price per token or provide equivalent licensing details for groundhog-1. Teams comparing options should verify the applicable terms and costs for the model and access route they plan to use.

The release gives teams two model profiles to evaluate for decisions that require calibrated probabilities, with long context and image input as stated capabilities. The benchmark and model descriptions can help shape an initial comparison, but teams still need to assess performance, cost and licensing against their own use case.

Why it matters
For teams

AI training for your team

kju.ai is built for teams: daily sessions, team analytics, and rollout support for your organization.

How this developed

  1. 21 November 2026

    Baseten releases two decision models with 128K context and image input

Sources

AI fluency, one session a day, built for your work.