GLM-5.2

GLM-5.2 (Hugging Face repo zai-org/GLM-5.2, API model id glm-5.2) is a Mixture-of-Experts model from Zhipu AI, distributed under the Z.ai brand as the flagship long-horizon coding/reasoning/agentic model in the GLM-5 family, succeeding GLM-5.1 (Hugging Face blog, accessed 2026-07-12). The models.dev registry (accessed 2026-07-12) records 2026-06-13 as the release date — the day GLM-5.2 first became available inside the GLM Coding Plan — while open weights and the inference-code repository followed on 2026-06-17 per Zhipu's own launch blog (Hugging Face blog). The weights are MIT-licensed and the serving code is Apache-2.0 (HF LICENSE; GitHub LICENSE) — a genuinely open-weight release by the vendor's own license files, though models.dev's registry flags it as closed-weight (see Key specs).

Key specs

Values as of 2026-07-12, cross-checked against models.dev and Zhipu/Z.ai's own documentation. Where the two disagree, this page follows the vendor's primary documentation (source hierarchy: vendor docs and license files over a third-party registry) and flags the gap rather than silently picking a number.

Spec	Value	Source
Context window	1,048,576 tokens (~1M), per Z.ai docs and the HF launch blog; models.dev's registry instead lists a round 1,000,000	Z.ai docs; HF blog; models.dev
Max output	131,072 tokens ("128K") for coding/agent tasks — matches models.dev's registry exactly; the HF blog separately reports 163,840 for reasoning-mode tasks, so the cap is task-mode dependent	Z.ai docs; HF blog; models.dev
Total parameters	753B total per the HF model card and launch blog; the GitHub README instead states 744B total with ~40B active — both are Zhipu-controlled primary channels and disagree with each other, unresolved as of 2026-07-12	HF model card; GitHub README
Modality	Text input, text output only — no vision or audio	Z.ai docs; models.dev
License	Open weights, MIT (Hugging Face); inference/serving code, Apache-2.0 (GitHub); models.dev's registry flags the model open_weights: false, contradicting the vendor's own license files	HF LICENSE; GitHub LICENSE; models.dev
Pricing, input	$1.40 / Mtok — matches models.dev exactly	Z.ai pricing docs; models.dev
Pricing, output	$4.40 / Mtok — matches models.dev exactly	Z.ai pricing docs; models.dev
Pricing, cached input	$0.26 / Mtok; cache storage listed as free "for a limited time" — matches models.dev exactly	Z.ai pricing docs; models.dev
Knowledge cutoff	Not stated on any primary Zhipu/Z.ai page found as of 2026-07-12 — an open gap, not a documented spec (see Limitations)	—
Release date	GLM Coding Plan access: 2026-06-13 (models.dev registry); open weights on Hugging Face/GitHub: 2026-06-17 (Zhipu launch blog)	models.dev; HF blog

Pricing and max-output figures cross-validate exactly against models.dev; context window, license classification, and the precise release-date field are the three points where the registry and vendor docs disagree, and this page resolves all three in the vendor's favor.

Architecture combines sparse MoE with DeepSeek-style Sparse Attention (DSA) and a mechanism Zhipu calls IndexShare, which reuses the same indexer across every four sparse-attention layers, cutting per-token FLOPs by 2.9x at 1M context length (HF blog; GitHub README). Reasoning effort is selectable per request: the GitHub README documents reasoning_effort as "max" (default) or "high", plus enable_thinking=false to disable thinking entirely; the Z.ai docs confirm the same toggle under a "type": "enabled" parameter (GitHub README; Z.ai docs).

Benchmark results

Benchmark	GLM-5.2	Comparison	Source	As of
SWE-bench Verified (Vals AI harness, independent)	82.8%	GPT-5.6 Sol 96.2%, Claude Fable 5 95.0%, Claude Opus 4.8 88.6%, Grok 4.5 86.6%, GPT-5.5 82.6% — same leaderboard	Vals AI	2026-07-12
SWE-bench Pro (Zhipu self-reported)	62.1%	GLM-5.1 58.4% — according to Zhipu	HF model card	2026-06-17
Terminal-Bench 2.1 (Zhipu self-reported)	81.0% on the Terminus-2 harness / 82.7% best-reported — two measurement conditions in the same HF model card, not a cross-artifact disagreement (the GitHub README's Terminal-Bench figure is the 81.0 Terminus-2 number)	GLM-5.1 63.5%, Claude Opus 4.8 85.0% (both on the Terminus-2 harness) — according to Zhipu	HF model card; GitHub README	2026-06-17
FrontierSWE	74.4%	Claude Opus 4.8 75.1% — according to Zhipu	HF blog	2026-06-17
SWE-Marathon	13.0%	Claude Opus 4.8 26.0% — according to Zhipu	HF blog	2026-06-17
GPQA-Diamond	91.2%	GLM-5.1 86.2%, Claude Opus 4.8 93.6% — according to Zhipu	HF model card	2026-06-17
AIME 2026	99.2%	GLM-5.1 95.3%, Claude Opus 4.8 95.7% — according to Zhipu	HF model card	2026-06-17
Humanity's Last Exam	40.5%	GLM-5.1 31%, Claude Opus 4.8 49.8% (asterisked in Zhipu's source, caveat unspecified) — according to Zhipu	HF model card	2026-06-17
MCP-Atlas (agentic tool use)	76.8%	GLM-5.1 71.8%, Claude Opus 4.8 77.8% — according to Zhipu (GLM-5.2 trails Opus by 1.0 pt)	HF model card	2026-06-17
Tool-Decathlon (agentic tool use)	48.2%	GLM-5.1 40.7%, Claude Opus 4.8 59.9% — according to Zhipu (GLM-5.2 trails Opus by 11.7 pt)	HF model card	2026-06-17

Full leaderboard: Vals AI — SWE-bench Verified leaderboard, as of 2026-07-12.

Reading notes for practitioners:

• On the one independent benchmark on this page, GLM-5.2 is a near-exact tie with GPT-5.5 — 82.8% vs 82.6% on Vals AI's SWE-bench Verified run — not with the other open-weight lab models in the same table: Kimi K2.7-Code (78.2%) and Kimi K2.6 (76.2%) both trail by several points (Vals AI, accessed 2026-07-12).
• Zhipu's Terminal-Bench 2.1 comparison numbers for other vendors' models don't match those vendors' own reporting. Zhipu's table credits Claude Opus 4.8 with 85.0% — above both Anthropic's own self-reported 74.6% (Anthropic announcement) and the independent Terminus-2/Claude-Code leaderboard's 78.9% best run (tbench.ai leaderboard). The same table's GLM-5.1 figure (63.5%) is likewise above that leaderboard's measured 58.7% for GLM-5.1. Treat every comparison-model score in Zhipu's Terminal-Bench table as Zhipu's own reproduction rather than that vendor's reported number, and extend the same caution to GLM-5.2's own 81.0–82.7% entry, which comes from the identical self-reported methodology.
• SWE-bench Verified and SWE-bench Pro are distinct benchmarks; the 82.8% and 62.1% rows above are not comparable to each other.
• The largest documented gap to Claude Opus 4.8 is on SWE-Marathon (13.0 vs 26.0, roughly half, by Zhipu's own reproduction) — the long-horizon agentic tasks this benchmark targets are exactly the profile GLM-5.2's 1M context and IndexShare efficiency work are aimed at, so this is the benchmark row most worth re-checking independently before relying on GLM-5.2 for that workload class.

Agentic behavior notes

• OpenAI/Anthropic-compatible API surface. GLM-5.2 supports streaming, function calling, structured/JSON output, context caching, and MCP tool integration, which is how third-party guides document wiring it into Claude Code and other harnesses as a swap-in backend via base-URL override — the endpoint/feature list itself is primary, but the specific "point Claude Code at it" setup steps come only from secondary how-to guides, not a Zhipu-published configuration (Z.ai docs, accessed 2026-07-12).
• Zhipu shipped its own agent harness, ZCode, around the GLM-5.2 release for autonomous coding-agent loops, with launch promotions to drive adoption — reported by a technology-desk outlet, not found on a primary Zhipu page in this pass (South China Morning Post, accessed 2026-07-12). See agent-harness for the general pattern.
• Self-hosting is a documented option, not just a license technicality. MIT-licensed weights plus multi-framework inference support (Transformers, vLLM, SGLang, xLLM, KTransformers, Unsloth, Ascend NPU) mean the model can run outside Zhipu's hosted API (GitHub README, accessed 2026-07-12).
• Cached input is priced separately at $0.26/Mtok against $1.40/Mtok base input — a lever for agent loops that re-send large system prompts or tool schemas on every turn (Z.ai pricing docs, accessed 2026-07-12).
• Zhipu's launch materials describe a two-stage anti-reward-hacking mechanism used during training (a rule-based filter plus an LLM judge) aimed at preventing benchmark/eval gaming (HF blog, accessed 2026-07-12) — a training-time claim, not an independently audited outcome.

Limitations

• Internal inconsistency in Zhipu's own primary artifacts. The total parameter count differs between two Zhipu-controlled channels — 753B per the Hugging Face model card/blog vs 744B (with ~40B active) per the GitHub README — a genuine gap in the vendor's own materials rather than an aggregator error (HF model card; GitHub README). The Terminal-Bench 81.0/82.7 spread is not such an inconsistency: both figures appear in the HF model card as two measurement conditions (Terminus-2 harness vs best-reported), and the GitHub README's Terminal-Bench number matches the 81.0 Terminus-2 figure.
• No independent verification for most benchmark claims, and Zhipu's own comparison numbers for other vendors run high. Every score in the table above except SWE-bench Verified is Zhipu self-reported. Where an independent check exists — the Terminal-Bench 2.1 leaderboard — Zhipu's own comparison table overstates both Claude Opus 4.8's score (85.0% vs a measured 78.9%) and GLM-5.1's own prior score (63.5% vs a measured 58.7%) (tbench.ai leaderboard, accessed 2026-07-12), which is reason to treat GLM-5.2's self-reported scores as an upper bound rather than a settled number.
• Trails Claude Opus 4.8 on several long-horizon/agentic tasks, by Zhipu's own comparison: SWE-Marathon (13.0 vs 26.0) and, on the Terminal-Bench methodology issue above notwithstanding, FrontierSWE (74.4 vs 75.1) — the gap narrows or closes on some non-agentic reasoning benchmarks (GPQA-Diamond, AIME 2026) but widens specifically on long-horizon coding/agent execution.
• Knowledge cutoff is undisclosed. No primary Zhipu/Z.ai page reviewed for this page states a training-data cutoff date; treat any specific date circulating elsewhere as unverified and do not rely on it for recency-sensitive tasks.
• Text-only modality. No documented vision or audio input/output, unlike several frontier competitors (Z.ai docs, accessed 2026-07-12).
• Dual-use commentary. Analysts have flagged that an unrestricted open-weight license lets GLM-5.2 be built into both defensive and offensive security tooling (large-scale automated code/vulnerability scanning), a framing raised in opinion/analysis coverage rather than a Zhipu-documented constraint (The AI Chronicle, accessed 2026-07-12).

Related

• Claude Opus 4.8 — Zhipu's own comparison table benchmarks GLM-5.2 directly against this model throughout the launch materials, though see the Terminal-Bench cross-check above before trusting those specific numbers.
• GPT-5.5 — GLM-5.2's closest peer on the one independently measured benchmark on this page (SWE-bench Verified, 82.8% vs 82.6%).
• Kimi K2.6 and Kimi K2.7-Code — the other prominent open-weight lab models on the same SWE-bench Verified leaderboard, both trailing GLM-5.2 by several points as of 2026-07-12.
• Context compaction — relevant to how a 1M-context, long-horizon coding model like GLM-5.2 is actually used across extended agent sessions rather than one-shot calls.