By ByteBot
Salesforce just shipped its first in-house reasoning model, and the headline numbers are real — but they need context before you build anything around them. Koa, post-trained on NVIDIA’s Nemotron-3-Super-120B using 30 years of synthetic CRM scenarios, is now in pilot with select Agentforce customers. If you’re building on Agentforce, here’s what actually changed.
What Koa Is
Koa starts with NVIDIA’s Nemotron-3-Super-120B — an open-weight Mixture-of-Experts Hybrid Mamba-Transformer. Salesforce applied SFT and GRPO reinforcement learning on top, using only synthetic data: no customer records, just simulated CRM workflows spanning 14-plus industries.
The training approach is what makes it interesting. Salesforce built a “simulation-to-reward” pipeline — each training scenario includes a persona, a task sequence, and a resolution reward tied to successful tool use. The model learned to pick the right Agentforce action, in the right order, not just generate a plausible text response. That’s a fundamentally different objective than what frontier generalists are trained on.
Salesforce holds the weights and runs inference inside its own infrastructure. If you’re on Agentforce, your CRM data stays inside Salesforce’s trust boundary when Koa handles the request — it doesn’t cross to Anthropic, OpenAI, or any external provider.
The Benchmark Numbers — All of Them
Salesforce’s “3x fewer errors” claim comes from its own CRM Bench, built from real tasks like updating an opportunity, routing a case, and scheduling a follow-up. On that benchmark: Koa (0.86) beats GPT-4.1 (0.81), roughly matches Claude Opus 4.8 (0.87), and trails GPT-5.5 (0.90).
On Tau2Bench, a third-party benchmark measuring end-to-end customer service conversations, Koa scores 69.41 — beating GPT-4.1 by nearly 15 points, but trailing Claude Opus 4.8 (74.00) and GPT-5.5 (83.99). Koa is also 11% more precise at calling the right tool action compared to GPT-4.1.
The honest read: Koa genuinely outperforms GPT-4.1 on CRM-specific tasks. It doesn’t outperform Claude Opus 4.8 or GPT-5.5. The “3x fewer errors” headline is from Salesforce’s own benchmark and doesn’t have independent validation yet. That’s not disqualifying — it’s just the information you need to make a real decision.
What Changes for Developers
If you’re building agents in Agent Script today, model selection via model_config already lets you route different subagents to different models. When Koa becomes available in your org, the pattern looks like this:
agent LeadProcessor {
subagent LeadQualification {
model_config {
model: "sfdc_ai__Koa" // CRM-tuned for tool precision
}
}
subagent ContractAnalysis {
model_config {
model: "sfdc_ai__DefaultGPT55" // General reasoning for edge cases
}
}
}Route structured CRM tasks — lead qualification, case routing, opportunity updates — to Koa. Route complex, open-ended reasoning to Claude or GPT-5.5. This isn’t theoretical; it’s what Agentforce’s model gateway is designed for. Note: Salesforce hasn’t published Koa’s API name yet. Get it from your account executive when you enter the pilot.
The Catch: It’s Not Available to Most Developers Yet
Koa is pilot-only today. General availability is targeted for winter 2026, US regions only. EU and APAC teams have no availability date — Salesforce has given no timeline for non-US data residency. Current pilot customers include Formula 1, Xero, UChicago Medicine, and 1-800Accountant.
If you want access now, contact your Salesforce account executive. If you’re evaluating Agentforce for an EU or APAC deployment, don’t plan around Koa for 2026.
Why This Matters Beyond the Salesforce Ecosystem
Koa is the enterprise domain-model thesis made concrete. Post-train an open-weight MoE model on constrained synthetic domain data using GRPO, and you can beat expensive frontier generalists on structured tasks at lower inference cost. Salesforce joins Mistral (Codestral for code), Mastercard (fraud detection), and Bloomberg (BloombergGPT for finance) in demonstrating that proprietary post-training on open foundations is a viable enterprise AI strategy.
CRM workflows are an ideal fit for this approach: fixed schema, known action space, deterministic success criteria. The model doesn’t need to write poetry or reason about physics — it needs to pick the right Salesforce action in a multi-step workflow. Koa is built exactly for that, and the benchmarks reflect it.
Independent validation is still missing. That’s the one thing worth watching. If third-party benchmarks confirm what Salesforce’s own data shows, Koa becomes a serious cost-reduction lever for Agentforce shops.
