Baseten doubles GLM-5.2 API performance and launches a lower-latency fast tier
philipkiely · x · 2026-07-26
Baseten says its GLM-5.2 serving stack now delivers more than 2× the launch-day performance, with peak speeds up to 280 tokens/s initially and a new benchmark showing much higher throughput and lower latency. The company also introduced a separate GLM-5.2-Fast API tuned for coding and agent workloads.
The optimization work included:
- scheduler tuning and bug fixes across the stack
- updated NVFP4 weights
- a revised speculative decoding profile
- latency-focused configuration changes for the fast API
- smaller max batch sizes and a shift away from Attention Data Parallelism toward tensor and expert parallelism
Both APIs run on NVIDIA B200 GPUs. The fast API trades throughput for latency and charges 50% higher input/output token prices. The post also notes that real-world performance depends heavily on traffic patterns and sequence lengths, and that more speculative-decoding improvements are planned.
Related event: Baseten Pushes GLM-5.2 to 280 tok/s(3 posts)→
More from Infra
- Anthropic's ~11 GW contracted compute estimated at ~42% NVIDIA GPUs, ~4.5 GW Google TPU — ns123abc · 2026-10-27
- EdgeRock-4B holds KV cache at 72 MiB across 256K context, 99.1% smaller than Qwen3.5-4B — alifcoder · 2026-10-27
- RTX 5080 price jumps from $1658 to $2318 in ten days on Amazon — Frizzy-MacDrizzle · 2026-10-27
- ComfyUI on 12GB VRAM: measured peaks show offloading turns 24s into 11 minutes — Impossible-Mail-4619 · 2026-10-27
- Neon rebuilds its docs search entirely on its own stack: Lakebase Search, Functions, AI Gateway — matei_zaharia · 2026-10-27
- PyTorch details 2.12x faster on-device streaming ASR on Arm SME2 via ExecuTorch — PyTorch · 2026-10-27