Post by DeepInfra

The gap between open-source AI and the closed-source frontier is closing faster than the consensus expects. A few months ago, the best open-source model on long-horizon coding trailed Claude Opus by 40+ points. Today Z.ai launched GLM-5.2 — built explicitly for long-horizon tasks — and it landed within 1% of Opus 4.8 on FrontierSWE, ahead of GPT-5.5 on PostTrainBench, and second only to Opus on SWE-Marathon. On Terminal-Bench 2.1 it jumped 17 points over its predecessor in a single release. Under the hood: 744B total / 40B active mixture-of-experts. A new architecture trick (IndexShare) cuts per-token FLOPs by 2.9× at full 1M context. High and Max thinking-effort levels let you trade latency for quality on harder coding tasks. MIT-licensed open weights. It's not an outlier. DeepSeek, Qwen, GLM, Kimi, MiMo — the labs shipping real open-weight models aren't chasing anymore. They're competing. DeepInfra exists for this moment. Day-zero hosting on every major release. OpenAI-compatible API across hundreds of models. Transparent pricing. GLM-5.2 went live this morning with Day-0 support. Full 1M context. $1.40 in / $4.40 out / $0.25 cached per 1M tokens. → https://lnkd.in/dGeGHAEG #AI #OpenSource #LLM