Baseten Bison-1 and Groundhog-1: Open Decision Models That Top the AI Leaderboard

Most AI models answer in paragraphs. Baseten’s newest models answer in probabilities. The inference company has released bison-1 and groundhog-1, two open-weight decision models built on Google’s Gemma 4 — and the bigger of the pair now sits at the top of the Artificial Analysis Decision Index, the benchmark that ranks this fast-growing class of AI.

Announced on November 20, 2026, the pair is Baseten’s second family of decision models, following mammoth-1 and elephant-1 from a month earlier. The headline numbers are striking: bison-1 scores 76 on the Decision Index v2.1, first among 19 models, with the best calibration on the index — and both models take eight times more context than their predecessors, 128,000 tokens instead of 16,000, while accepting images as well as text.

For teams building AI agents, fraud checks, content moderation and other systems that must choose rather than chat, that combination — open weights, long context and calibrated odds — is exactly the point.

What are bison-1 and groundhog-1?

Bison-1 is the flagship: a dense 31-billion-parameter model post-trained from Gemma 4 31B. Groundhog-1 is its cheaper sibling, a 26-billion-parameter mixture-of-experts model based on Gemma 4 26B-A4B, with only about 4 billion parameters active per token.

Both are decision models, not chatbots. Instead of generating free-form text, they return calibrated probabilities for predefined decisions — approve or decline, safe or unsafe, this category or that one. They were trained on 6.71 million decision examples spanning core decision skills, safety and abstention, judging and expert domains, documents and long-context tasks.

The pair joins Baseten’s existing lineup: mammoth-1 (27B), elephant-1 (9B) and joey-1 (2B), all built on Qwen backbones. The new models are the company’s first on what it calls an American backbone — Google’s Gemma 4 — a switch it frames as an investment in American technical strength and open models.

What is a decision model, anyway?

A developer studying an AI decision dashboard with probability gauges, a branching decision tree and confidence bars

Decision models are one of the quieter revolutions in applied AI. A general-purpose chatbot is brilliant at open-ended conversation but awkward when a product needs a yes-or-no answer it can trust, log and act on. A decision model is trained for exactly that job: given a prompt, a document or an image, it returns the probability of each possible outcome.

That matters because probabilities can be thresholded, audited and combined. A fraud team can auto-approve everything above 95 percent confidence and send the rest to a human. A moderation system can escalate only the borderline cases. A coding agent can check its own work before committing it. The model does not need to explain itself in prose — it needs to be right, and to know how sure it is.

Baseten’s bet is that this narrow job deserves its own models, benchmarks and pricing, separate from the chat models that dominate the headlines. The Decision Index from Artificial Analysis exists precisely to rank them.

Bison-1 takes the top spot

On that index, bison-1 leads the field. Its score of 76 on Decision Index v2.1 puts it one point ahead of Liquid AI’s d1.2-ultra at 75, and five points ahead of Baseten’s own mammoth-1 at 71. Nineteen decision models were evaluated in this version of the index.

Just as important as the headline score is the calibration behind it. Bison-1 posts a Brier score of 0.14 — the lowest on the index, against 0.15 for d1.2-ultra — meaning its stated probabilities line up unusually well with reality. A model that says it is 80 percent sure and is right 80 percent of the time is far more useful in production than one that is equally accurate but overconfident.

On long-context work, the pair also looks strong: on LongMemEval S, the two models score 87.1 and 79.3. Groundhog-1 has not yet received an official Decision Index score, though Baseten expects it to match mammoth-1.

The timing is notable, too. The open-weights world has had a busy week — the Inkling 2 Mini open-weights model arrived the same day, built to run on a laptop — and decision models are now getting the same open treatment as chat models.

128K context and image input change the game

The biggest practical upgrade over mammoth-1 and elephant-1 is context. The new pair accepts 131,072 tokens — 128k — eight times the 16k of the previous generation. Baseten’s own pitch is that teams can now feed in full policies or whole conversation histories without selecting excerpts or summarising first.

Both models also accept image input alongside text, so a decision can be based on a screenshot, a scanned document or a photo, not just words. That opens up use cases like document classification, visual quality checks and multimodal moderation that the earlier text-only pair could not handle.

There is a cost to the bigger models, of course. The open BF16 weights run to roughly 62 GB for bison-1 and 52 GB for groundhog-1, so self-hosting them is a serious GPU commitment. Most teams will likely use the hosted API instead.

The American backbone bet

Two glowing orbs in blue and orange connected by a network of light threads, representing an AI model backbone

The switch from Qwen to Gemma 4 is the most political detail of the release — and Baseten is not hiding it. Mammoth-1 and elephant-1 were built on Qwen, a Chinese base model. Bison-1 and groundhog-1 run on Google’s Gemma 4, and the company frames the move as giving teams a reason to choose American models on performance rather than preference alone.

The framing matters for buyers in government, regulated industries and sovereignty-sensitive markets, where the provenance of a base model is increasingly part of the procurement conversation. But it is also a performance claim: Baseten is arguing that the American backbone is now the better one, not just the safer choice.

Licensing follows the base model. Google’s Gemma 4 carries Apache License 2.0 terms, and Baseten’s modifications are released under the MIT License, with use subject to both.

Pricing and how to try them

Pricing is deliberately boring — in a good way. Bison-1 costs $0.10 per million tokens, exactly what mammoth-1 cost, and groundhog-1 costs $0.035 per million tokens, exactly what elephant-1 cost. Existing customers can move to the new models without repricing their workloads, since no output tokens are produced and the price is an input price.

Both models are live now on Baseten Model APIs under the names decision-models/baseten-bison-latest and decision-models/baseten-groundhog-latest, using the same API as the older decision models — switching means changing one model name. The open weights are also available on Hugging Face for teams that want to run the models themselves.

That drop-in approach mirrors a wider pattern in AI infrastructure this week: OpenAI’s GPT-Live-1.1 voice model also shipped with a pricing change designed to make adoption painless, cutting idle standby cost for voice agents.

What to watch next

Three things will decide whether bison-1 and groundhog-1 matter beyond the benchmark. First, groundhog-1’s official Decision Index score — if it lands near mammoth-1 as Baseten expects, the cheaper model becomes the value pick of the family. Second, whether developers actually use the 128k context and image input in production, or whether most decision workloads stay short and text-only. And third, whether the “American backbone” framing pulls in the government and regulated-industry customers Baseten is clearly courting.

The broader question is whether decision models become a standard layer of the AI stack — a small, cheap, trustworthy model sitting behind every agent, checking every consequential choice — or remain a niche for specialists. Baseten, now valued at $13 billion after a $1.5 billion funding round and running inference for companies like Cursor and Abridge, is betting on the former.

For now, the leaderboard says the bet is off to a strong start: a 31-billion-parameter open model, built on an American backbone, sitting at No. 1 — and a 4-billion-active-parameter sibling waiting in the wings at a third of the price.

In related AI news this week, Nothing’s first AI device, Oppo’s Heartball AI wearable and OpenAI’s reported AI audio headset show how fast AI is moving into hardware.

Sources: Baseten’s launch post, MarkTechPost’s breakdown of the release, and Unite.AI’s report on the new models.

Leave a Comment

Your email address will not be published. Required fields are marked *

Scroll to Top