Clay Just Made Open-Weight Models Production-Grade for GTM Research, Not Just Content Generation

Clay added two open-weight models, Kimi K2.6 and GLM 5.2, to Claygent, its AI agent for GTM research tasks like account research, lead qualification, scoring, and outbound copy — available in the model dropdown for any Claygent column at a fraction of closed-model cost, with Fireworks AI as the inference partner. The company has offered open-source models for about a year, since DeepSeek R1, but restricted them to content generation because agentic research demands multi-step planning, browsing, and reasoning earlier open models couldn’t sustain. Clay’s internal testing says the newest generation clears that bar. The economics matter more than the model swap itself: Claygent runs are long and token-heavy across tables with tens of thousands of rows, so cheaper inference means researching an entire TAM instead of a sample, qualifying every inbound lead instead of a subset, or re-scoring accounts weekly instead of quarterly.

The architectural leap Clay is betting on shows up in Moonshot AI’s newly released Kimi K3, a 2.8-trillion-parameter Mixture-of-Experts model that shipped its API on July 16 with full open weights due July 27. K3’s quantization-aware training stores weights in MXFP4, shrinking storage to roughly 1.4TB versus 5.6TB for an FP16 equivalent — the kind of efficiency gain that makes self-hosting frontier-scale reasoning newly practical for a vendor like Clay instead of requiring proprietary inference infrastructure from a closed lab.

The Ceiling Method read: this is a supply-side shift with a direct effect on what “AI research” costs a rep or RevOps team, not a capability story that shows up in a demo. If your outbound tooling gets cheap enough to research a full TAM instead of a sampled list, the differentiator stops being “we used AI to research this account” — every competitor using the same tooling gets the same coverage. What’s left to differentiate on is what a rep does with a fully-researched account that a script can’t: the specific commercial insight that makes the first call worth taking.