hpcaitech
GLM 5.2 is now available via a unified Model API
New API Models Launched
Model API Performance
We’ve been benchmarking a few models on our API platform and got some interesting performance numbers: - MiniMax M2.5 → 0.118s time-to-first-token, 103 tokens/sec - GLM 5.1 → 120 tokens/sec throughput - Kimi K2.5 → 0.643s TTFT, 69 tokens/sec - All models → ~99.9% request success rate The latency difference is especially noticeable, ~0.1s TTFT feels almost instant in interactive...
The case for Model-as-a-Service over self-managed inference
The operational overhead of self-hosting model inference (vLLM setup, networking, scaling) is significant for small teams. A growing category of "MaaS" platforms abstracts this, so i've been testing hpc-ai.com for this purpose & its been working well. Built on Colossal-AI (~41k GitHub stars), & external endpoints with minimal config. Curious what HN thinks: right abstraction, or too...
MaaS Updates
Model APIs
What's the Best LLM for Coding in 2026
Calling all who run inference in models
We've been thinking a lot about inference infrastructure recently, and it seems like the challenges are very different from training. Training tends to be compute-heavy but predictable, while inference introduces things like: - latency constraints - dynamic batching - unpredictable traffic patterns - model versioning in production - GPU utilization issues at smaller batch sizes For people...
Earn Compute Credits
HPC-AI explains embodied AI
AI Startups and GPUs
Startups can now train trillion-parameter models without owning a cluster. For years, scaling AI was an "arms race" only big tech could afford. GPU scarcity and cluster management meant smaller teams were locked out. Now, on-demand GPU access lets startups and research labs experiment with huge models without massive CapEx, spin up temporary clusters for heavy workloads, & pay only for what...