GLM-5.2

What GLM-5.2 is

GLM-5.2 is a large language model released by Zhipu on 16 June 2026 under the MIT License, a permissive license that in general allows reuse and modification of the weights.^[2][1] Its model card describes a 1M-token context window, raised from 200K in the preceding generation.^[2][1] Zhipu's blog post introduced the release under the title "GLM-5.2: Built for Long-Horizon Tasks".^[1]

The model card gives its base model as 753B parameters and notes that the same base was later used by GLM-5.3.^[1][2] Wikipedia lists the base at 743B parameters while agreeing that GLM-5.3 reused it.^[2] Weights are published in BF16 and F32.^[1]

Architecture and efficiency

The main efficiency change is IndexShare, which reuses a single sparse-attention indexer across every four layers. At a 1M-token context this cuts per-token floating-point operations by 2.9 times.^[1] Sparse attention needs an indexer that decides which parts of a long context a layer should look at; sharing one indexer among several layers reduces the cost of that bookkeeping.

A revised multi-token-prediction layer raises the acceptance length of speculative decoding by up to 20%.^[1] Speculative decoding speeds up generation by having a cheap draft propose several tokens that the main model then checks in one pass, so a longer accepted run means more of the draft survives verification. The model also offers several thinking-effort levels, which trade answer quality against latency.^[1]

Benchmark results

The model card compares GLM-5.2 with GLM-5.1 on several benchmarks. On HLE it scores 40.5 against 31, and 54.7 against 52.3 when tools are allowed. On GPQA-Diamond it scores 91.2 against 86.2, and on AIME 2026 it scores 99.2 against 95.3.^[1]

On coding tasks the gaps are larger in places: SWE-bench Pro 62.1 against 58.4, Terminal Bench 2.1 at 81.0 against 63.5, and DeepSWE at 46.2 against 18.^[1] The card notes that some evaluations — including SWE-bench Pro, NL2Repo and DeepSWE — were run with a 400K context window rather than the full 1M.^[1]

The card's comparison table also places the model against systems from other developers: it lists Claude Opus 4.8 at 69.2 on SWE-bench Pro, above GLM-5.2's 62.1.^[1]

Serving

According to the model card, serving GLM-5.2 on SGLang requires version v0.5.13.post1 or newer, and on vLLM version v0.23.0 or newer.^[1]

Place in the line

GLM-5.2 was succeeded by GLM-5.3 on 14 August 2026, which used the same base model.^[2] The GLM-5.2 model card notes that a newer version, zai-org/GLM-5.3-BF16, is available.^[1] The pattern — a base model kept in service across two released generations while attention, prediction and serving details change — is what makes GLM-5.2 useful to read alongside its successor.