GLM-5.2 Official Release Notes and Complete Benchmark Table (Z.ai Blog)

Core content summary

GLM-5.2 is Z.ai's flagship model for "long-horizon tasks." Its key selling points are a genuinely usable 1M-token context, flexible reasoning modes, and the MIT open-source license (with no regional restrictions). According to the official account, it delivers a substantial improvement over GLM-5.1 on long-horizon tasks and is the first model to "reliably withstand engineering pressure" with a 1M context.

Release highlights

• 1M context: Training on a 1M context was substantially expanded for coding-agent scenarios (large-scale implementation, automated research, performance optimization, and complex debugging), emphasizing that "a 1M context must be engineering-usable, not merely able to accept more tokens."

• IndexShare architecture: One indexer is shared by every four layers of sparse attention, reducing FLOPs per token by 2.9× at a 1M context length; improvements to the MTP (multi-token prediction) layer are used for speculative decoding, increasing acceptance length by up to 20% (ablation: baseline 4.56 → +IndexShare+KV Share 5.10 → +Rejection Sampling 5.29 → +End-to-end TV Loss 5.47).

• Inference serving: LayerSplit provides fine-grained memory management and parallelism; long-context kernels and CPU-side cache management are optimized, with throughput advantages growing as the context gets longer.

• Agentic RL (slime framework): Parallel OPD training combines 10+ expert models, with the entire OPD training taking about two days; critic-based PPO (single rollout, token-level advantage) is introduced to support compaction; coding RL adds two-stage anti-cheating (rule-based filter + LLM judge), intercepting hack calls online and returning fake data so the rollout can continue.

• Access: GLM Coding Plan (model name GLM-5.2; use GLM-5.2[1m] in Claude Code to enable a 1M context); Z.ai chat; open-source weights on HuggingFace / ModelScope (supporting transformers, vLLM, SGLang, xLLM, and ktransformers).

Complete benchmark table (vendor-reported, 2026-06-16)

(* Full-set scores.)

Three long-horizon benchmarks (official figures)

• FrontierSWE (Proximal evaluation, 1M context + Max mode + 128K output): 74.4, only about 1% behind Opus 4.8 (75.1), about 1% ahead of GPT-5.5 (72.6), and about 11% ahead of Opus 4.7.

• PostTrainBench (each agent gets an H100, measuring how much it can improve a small model through post-training): 34.3, second only to Opus 4.8 (37.2).

• SWE-Marathon (ultra-long-horizon tasks: writing a compiler, optimizing kernels, and developing production-grade services): 13.0, about 13% behind Opus 4.8 (26.0) (as stated in the original), and still the top open-source model.

Official evaluation-method footnotes (key to reproducibility)

• HLE and other reasoning tasks: temperature=1.0, top_p=0.95, maximum generation length 163,840; the text-only subset is reported by default.

• SWE-bench Pro: OpenHands + custom instruction prompt, temperature=1, top_p=1, max_new_tokens=32k, 400K context.

• DeepSWE: official pier evaluation framework + mini-swe-agent, temperature=1.0, top_p=1.0, timeout=2h, 400K context, isolated container (2 CPUs, 8GB RAM, no network).

• Terminal-Bench 2.1: Terminus-2 framework (parser=json, timeout=4h, max_new_tokens=48k, max_episodes=500, 256K context, 4 CPU / 8GB RAM limit); the Claude Code version uses a transparent proxy to raise max_new_tokens to 128k, removes the wall-clock limit, and averages five runs.

• MCP-Atlas: think mode, a public subset of 500 tasks, a 10-minute timeout per task, and Gemini-3.0-Pro as the judge.

• Usage note: Different models may use different harnesses, so cross-row comparisons should be made cautiously.

Quotas and costs (Coding Plan figures)

• Quota consumption is 3× during peak periods and 2× off-peak; under the limited-time promotion (through the end of September), off-peak usage is counted as 1×. Peak hours are 14:00–18:00 (UTC+8) every day.

• The official recommendation for coding tasks is Max mode.

Key quotations from the original

"Supporting long-horizon tasks starts with making long context engineering-usable: the model must maintain quality acros… This is a necessary excerpt; read the original source for full context.

"A 1M context is easy to claim, but much harder to keep reliable under real engineering pressure."

"Across all three benchmarks, GLM-5.2 is the highest-ranked open-source model, showing that its 1M context has translate… This is a necessary excerpt; read the original source for full context.

Scope and limitations

• All benchmark scores are vendor-reported, including results run by third-party evaluation organizations (Proximal / PostTrainBench / Abundant AI); when comparing with closed-source models, note that the models may use different harnesses, contexts, and modes.

• When citing external scores, it is recommended to provide the evaluation configurations from the official footnotes as well; the 5.2 baselines published officially when GLM-5.3 was released (such as Terminal-Bench 3.0 4.6 and DeepSWE v1.1 46.2) use different benchmarks from the 2.x versions in this article, so distinguish the versions carefully.