Baseten vs Khoj

The catalog turns over constantly — DeepSeek V4 Pro 0813, Inkling and Inkling Small, Kimi K3, GLM 5.2 Fast in, older GLM and Kimi entries out — all reachable through the same OpenAI-compatible endpoint with dedicated deployments for larger workloads. Two releases break that pattern: Baseten for Model Labs packages the serving stack as infrastructure a lab can adopt instead of building its own, and the Fast tier debuts with GLM 5.2 Fast, identical weights on dedicated capacity tuned for sustained per-user throughput. The platform work underneath is now mostly enterprise plumbing — org-scoped key administration, programmatic logs and metrics, GPU usage for admins, and now runtime OIDC so deployments reach cloud providers without stored long-lived credentials.

Baseten is working both sides of the market at once. Toward developers, model choice is being commoditised into interchangeable catalog entries while serving characteristics become the thing actually priced. Toward labs, the pitch is that distribution and serving are someone else's problem. Both converge on the same position: whoever owns the endpoint owns the relationship, regardless of who trained the weights. The recent credential and observability work is the unglamorous prerequisite for the accounts that position requires.

Expect the Fast tier to expand beyond GLM 5.2 to the models agentic workloads lean on hardest, and the deprecation cadence to keep thinning older catalog entries as newer ones land. Whether Model Labs attracts a named lab publicly is the thing these entries cannot yet show.