GLM-5.2 (flexai)

The GLM-5.2 release brief records a June 13, 2026 launch that shipped first to paying GLM Coding Plan customers, with MIT open weights following days later on Hugging Face as zai-org/GLM-5.2 without regional restrictions. The model is a 753B total / roughly 40B active mixture-of-experts with a 1,000,000-token context and 128,000 max output tokens, list priced near $1.40 per million input and $4.40 per million output tokens. Access spans the Z.ai API, the Coding Plan at roughly $10, $30, and $80 per month, Hugging Face weights, and self-hosting through transformers, vLLM, SGLang, KTransformers, and Ascend NPU, which decouples deployments from NVIDIA supply. It drops into Claude Code or Cline with a configuration change, and the brief ranks GLM-5.2 first among open-weight models on the Artificial Analysis Intelligence Index at 51.

Z.ai published the GLM 5.2 benchmark scorecard it withheld at the June 13, 2026 launch, reporting 62.1 on SWE-bench Pro versus GPT-5.5 at 58.6 and GLM 5.1 at 58.4, plus 81.0 on Terminal-Bench 2.1, 76.8 on MCP-Atlas, and 91.2 on GPQA-Diamond. The scores now live in the model card and a long-horizon blog post. On June 17, 2026, Z.ai uploaded the full MIT-licensed weights for zai-org/GLM-5.2 and a zai-org/GLM-5.2-FP8 quantized build to HuggingFace, earlier than the promised following week, with 24 community quantization variants for llama.cpp, LM Studio, Jan, and Ollama. The model also works with Claude Code, the Claude Agent SDK, Cursor, and the Vercel AI Gateway.

Z.ai launched GLM 5.2 on June 13, 2026, making it live across all GLM Coding Plan tiers (Lite, Pro, Max, Team) immediately while the standalone API, the Z.ai chatbot, the MIT-licensed open weights, and the technical report were scheduled for the following week. The headline spec is a 1,000,000-token context window using the model ID glm-5.2[1m], with maximum output capped at 131,072 tokens. No benchmark numbers were published at launch, so claims about coding power and long-horizon work remained unverified. An August 14, 2026 update notes that GLM-5.3 has since shipped, that the GLM Coding Plan now auto-routes GLM-5.2 and GLM-5.1 requests to 5.3 with no documented opt-out, and that the plan shifted to a credit system with per-model multipliers.

The GLM-5.2 tracking entry lists the model as open-weight, released on June 16, 2026, seventy days after GLM-5.1, with a 744B parameter mixture-of-experts design and a one-million-token context window. Benchmark results on the page cover BullshitBench v2, ProgramBench, SWE-Bench Pro, DeepSWE 1.1, Next.js Evals, and Frontier-Bench v0.1. Pricing is listed at $1.40 per million input tokens, $0.26 cached input, and $4.40 output, with rates verified against docs.z.ai on October 19, 2026. Reasoning output is supported, and the entry breaks availability down by serving provider, region, and service tier.

Z.ai released GLM-5.2 on 16 June 2026 as an MIT-licensed open-weight engineering model with roughly 750 billion total parameters and about 40 billion active per token, using a one-million-token context window. Z.ai says IndexShare cuts per-token FLOPs by 2.9 times at one million tokens, while its speculative-decoding work raises acceptance length by 20 percent. Vendor-reported results include Terminal-Bench 2.1 at 81.0 (up from 63.5 on GLM-5.1), SWE-bench Pro at 62.1, FrontierSWE at 74.4, and PostTrainBench at 34.3, while the harder SWE-Marathon reached only 13.0 against 26.0 for Opus 4.8. Artificial Analysis scored the model 51 on its Intelligence Index at about 1524 Elo on GDPval-AA v2, generating roughly 193 tokens per second.