Baseten

Baseten's Writer case documents private deployments of Palmyra-Med-70B and Palmyra-Fin-70B using TensorRT-LLM engines, with Writer CTO Waseem Alshikh describing production model-serving support.

Baseten customer deployments include Notion, OpenEvidence, World Labs, Poolside and other AI-native companies; Parallel reported 3x higher throughput and 2x latency improvement using Baseten inference infrastructure.

Baseten customer deployments include Notion, OpenEvidence, World Labs, Poolside and other AI-native companies; Parallel reported 3x higher throughput and 2x latency improvement using Baseten inference infrastructure.

Baseten describes a hybrid speculative-decoding mechanism that combines suffix-automaton lookup with MTP or EAGLE draft generation, choosing lookup when matches exceed a threshold. CUDA updates overlap KV-cache prefill and integrate with TensorRT, supplying a concrete inference optimization distinct from its existing general serving and distribution portfolio evidence.

Baseten has expanded from model serving into performance-optimized inference, training-to-deployment workflows, and model-lab distribution infrastructure including Frontier Gateway and Baseten for Model Labs.

Baseten has expanded from model serving into performance-optimized inference, training-to-deployment workflows, and model-lab distribution infrastructure including Frontier Gateway and Baseten for Model Labs.

Baseten announced a completed $150 million Series D on September 5, 2025 led by BOND, with Conviction, CapitalG and existing investors participating.

Baseten supports high-performance production inference and closed-weight model distribution for customers such as Poolside, Subconscious, Trajectory and WRITER, indicating strong production maturity without hyperscale breadth.

Baseten supports high-performance production inference and closed-weight model distribution for customers such as Poolside, Subconscious, Trajectory and WRITER, indicating strong production maturity without hyperscale breadth.

Baseten published an open-source suffix-automaton accelerator integrated with TensorRT-LLM: host construction overlaps GPU KV prefill, state transfers to the device, and C++/CUDA-compatible structures update during decoding. These concrete GPU runtime and data-structure skills are transferable beyond model serving.

Baseten work spans inference optimization, model serving, deployment, training loops, developer infrastructure and model distribution, creating broadly portable production-AI skills.

Baseten work spans inference optimization, model serving, deployment, training loops, developer infrastructure and model distribution, creating broadly portable production-AI skills.