GLM-5.2

Overview

GLM-5.2 is Zhipu AI's open-weight flagship, announced on June 16, 2026 — the date carried by Zhipu's own release notes, which credit the model with "1M lossless context, significantly improving long-horizon task capabilities and reducing context drift and goal forgetting in complex tasks." The weights landed on Hugging Face the same day as zai-org/GLM-5.2. Zhipu describes it as "a flagship model built for the era of long-horizon tasks," and the model card is unambiguous about what changed: it represents "a substantial leap in long-horizon task capability over its predecessor GLM-5.1."

The model is 753 billion parameters running a sparse Mixture-of-Experts architecture that the config identifies as glm_moe_dsa — a MoE stack whose attention layers are themselves sparse, using an IndexShare mechanism. Sparsity in both the feed-forward and attention paths is what pays for the headline capability: a 1M token context window that Zhipu claims is not a marketing figure but a "truly usable 1M-token context," tested against "project-scale engineering context."

It ships under the MIT License with no regional limits. Among models scoring above 60 on SWE-bench Pro, that licensing posture is rare.

Capabilities

• Long-horizon agentic work: 81.0 on Terminal-Bench 2.1 (Terminus-2), against 85.0 for Claude Opus 4.8 in Zhipu's own comparison — the closest an openly licensed model has come on this benchmark.
• Real-world software engineering: 62.1 on SWE-bench Pro, up from GLM-5.1's 58.4.
• Frontier mathematics: 99.2 on AIME 2026.
• Graduate-level science: 91.2 on GPQA-Diamond.
• Tool-augmented reasoning: 40.5 on Humanity's Last Exam, rising to 54.7 with tools — up from 31 for GLM-5.1.
• Production API surface: thinking mode with multiple options, streaming, function calling, context caching, structured JSON output, and MCP tool integration.

Technical Specifications

• API model ID: glm-5.2
• Hugging Face repository: zai-org/GLM-5.2
• Total parameters: 753B
• Architecture: glm_moe_dsa — sparse Mixture-of-Experts transformer with sparse attention layers using IndexShare
• Context window: 1M tokens
• Max output: 128K tokens per the z.ai documentation; the Hugging Face card cites generation lengths up to 163,840 tokens for reasoning runs
• License: MIT, no regional limits
• Serving: vLLM and SGLang

Zhipu has not published an activated-parameter count, a training-token count, or a knowledge cutoff for GLM-5.2. Those fields are deliberately absent from this page rather than estimated.

Relationship to GLM-5.1

GLM-5.1 preceded it — its weights first appeared on Hugging Face in early April 2026, not on the mid-April date that circulated widely. The two models share an identical rate card on z.ai. The difference is capability, concentrated in the places long-horizon agents actually break: sustained coding sessions, tool use, and reasoning that has to survive many turns.

Use Cases

• Project-scale engineering agents: A 1M context plus 81.0 on Terminal-Bench 2.1 targets agents that hold an entire service in context and work through it over hours.
• Whole-repository refactors and migrations: The context window removes most of the retrieval scaffolding that smaller-window models require.
• Research assistance in science and mathematics: 91.2 GPQA-Diamond and 99.2 AIME 2026 place it in the frontier band on both.
• Tool-using research agents: The HLE jump from 40.5 to 54.7 with tools indicates the model uses retrieval well rather than pretending to know.
• Private and regulated deployment: MIT weights with no regional restrictions make on-premises and air-gapped hosting straightforward.
• Cost-controlled long context: A $0.26 per 1M cached-input rate makes repeated million-token prompts economically defensible.

Performance / Benchmarks

Reported on Zhipu's GLM-5.2 model card.

For reference, the same card reports Claude Opus 4.8 at 85.0 on Terminal-Bench 2.1 — a comparison Zhipu chose to publish, and a 4-point gap it does not hide.

These are vendor-reported figures. Treat them as an upper bound and validate on your own evaluation set.

Limitations

• No published active-parameter count: Serving cost and latency are hard to model in advance because Zhipu has not disclosed how many of the 753B parameters activate per token.
• Still behind the closed frontier on agentic tasks: 81.0 versus 85.0 on Terminal-Bench 2.1 is a real gap on exactly the workload the model is marketed for.
• Self-hosting scale: 753B parameters is a multi-node deployment. This is not a model you run on a workstation.
• No published knowledge cutoff: Ground time-sensitive queries with retrieval or tool use.
• Conflicting max-output figures: The API docs say 128K; the model card cites reasoning generations up to 163,840 tokens. Verify against the endpoint you use.
• Weights, not full openness: MIT covers the weights. Training data and pipeline are not published.
• Vendor-reported benchmarks: Independent replication of the headline scores is limited.

Pricing & Access

From the z.ai pricing page, per 1M tokens in USD.

Cached input storage is currently listed as free for a limited time.

Access options:

• z.ai API — the primary managed endpoint
• Hugging Face — MIT-licensed weights at zai-org/GLM-5.2
• ModelScope — mirror for faster downloads in China
• vLLM / SGLang — high-performance local serving

Ecosystem & Tools

• GLM-5.2 documentation — thinking modes, function calling, context caching, structured output, MCP
• GLM-5.2 on Hugging Face — weights and the benchmark table
• z.ai — platform, console, and chat interface
• vLLM — production serving with tensor and pipeline parallelism
• SGLang — structured generation and high-throughput serving
• MCP — native Model Context Protocol tool integration

Community & Resources

• GLM-5.2 guide
• z.ai release notes — dated model announcements, including the June 16, 2026 GLM-5.2 entry
• Pricing
• zai-org on Hugging Face
• Compare with DeepSeek V4, Kimi K2.6, Tencent Hy3, and Ling-2.6-1T