Not tested by us yet. Koa Decisions has no public API and the Hugging Face repository we found is private, so our harness cannot call it. Everything below comes from Salesforce's announcement, the Unite.AI write-up and the Hugging Face model card. We'll update this post when a developer API or open weights let us run it on our own prompts.
Koa Decisions is Salesforce's decision model: you give it a request and a list of options you define, and it returns one choice, a yes/no, or a score instead of a paragraph. It accepts text, images and audio, it is built on Google's open-weight Gemma, and today it is only available inside Salesforce products, mainly for routing in Agentforce. Salesforce publishes no price.
| Koa Decisions | |
|---|---|
| Maker | Salesforce |
| Announced | 2026-11-06 |
| Type | Decision model (single-step pick, yes/no or score) |
| Base | Google Gemma, with Salesforce's own training techniques |
| Inputs | Text, images, audio |
| Predecessor | HyperClassifier |
| Access | Agentforce routing; free demo with a Trailhead account |
| Planned | Flows, Apex, prompt templates, developer API |
| Price | Not published |
What it is and what changed
A decision model is built for one job: classify, route or score. Because the answer is a short code, not generated text, it can answer in a fraction of the time a chat model needs. Salesforce describes Koa Decisions as the next version of HyperClassifier and says new categories can be added without retraining, so you change the option list and use it right away. The new part in this release is multimodal input: a request can carry an image or an audio clip, not only text.
For a similar open-weight release, see our explainer on Kakao's Kanana-2-v-decide, which returns probabilities for each option.
Vendor-reported numbers
None of these are ours, and none have been independently reproduced.
| Claim | Figure | Source |
|---|---|---|
| Routing latency, decision model vs general LLM | About 26 ms vs about 1.4 s | Salesforce announcement |
| Accuracy on a production intent-routing benchmark | 85.5% | Preprint, as reported by Unite.AI |
| Median speed vs frontier models | Roughly 7.5x faster | Same preprint, authors' own finding |
Read the latency row carefully. The Salesforce page attributes 26 ms to its Agentforce routing tests, while Unite.AI says the 26 ms and 1,446 ms figures describe the HyperClassifier predecessor and "do not establish" Koa Decisions' own performance. Treat the exact speedup as unconfirmed until a hosted version can be timed.
The model card's own checks are small smoke tests: 8 of 8 on image-only tasks, 8 of 8 on mixed image and audio, 12 of 12 on audio-only and 8 of 8 on text-only. Those sets are tiny, so they show the model runs, not how accurate it is.
Access and caveats
You can try the model in a free demo with a Trailhead account, where you edit the request and categories and see confidence scores. Outside Salesforce there is no hosted endpoint yet, and the card says the runtime needs a bundled vLLM extension, validated on a single NVIDIA B300.
The card also lists what is unproven. The audio projector was transferred onto a different fine-tuned backbone rather than trained jointly with it. Open-ended transcription, noisy or multilingual audio, complex image reasoning, longer contexts and video all need separate evaluation. The card says use is subject to Gemma and source-component license terms. The announcement links a research paper, and the card and the announcement describe the base model slightly differently (Gemma versus Gemma 4 with an Ultravox audio component), so check the license before building on it.
Who should care
If you already run Agentforce, Koa Decisions is the routing model working behind your agents, and the multimodal input matters for cases like routing a photo of a damaged product. If you are choosing a decision model for your own stack, wait: there is no price, no API and no way to benchmark it against an LLM classifier on your data. Until then, our guide to the best LLMs for structured output in 2026 covers the chat-model alternative for classification.