GLM-5.2 on Run BiOS

GLM-5.2 is the flagship open-weight language model from Z.ai (formerly Zhipu AI), released in June 2026 as the successor to GLM-5.1. Z.ai positions GLM-5.2 as a model "built for long-horizon tasks": it marks a substantial leap in long-horizon task capability over GLM-5.1 and, for the first time in the series, delivers that capability on a solid 1M-token context window. According to the Z.ai announcement, GLM-5.2 is the highest-ranked open-source model on all three of the long-horizon coding benchmarks it reports, showing that its 1M context translates into practical long-horizon delivery capability.

Headline capabilities highlighted by Z.ai:

• Solid 1M context — a 1M-token context (1,048,576 tokens) that stably sustains long-horizon work.
• Advanced coding with flexible effort — stronger coding capabilities with multiple thinking effort levels to balance performance and latency.
• Improved architecture — the new IndexShare technique reuses the same indexer across every four sparse attention layers, reducing per-token FLOPs by 2.9× at a 1M context length; the MTP (multi-token prediction) layer used for speculative decoding is also improved, increasing acceptance length by up to 20%.
• Pure open — an MIT open-source license, described by Z.ai as "no regional limits, technical access without borders."

Architecture

GLM-5.2 is a mixture-of-experts (MoE) model with sparse attention. Published specifications:

• Scale: 744B total parameters with 40B active per token (744B-A40B); BF16 weights (an FP8 variant is also published).
• Attention: DeepSeek Sparse Attention (DSA) with the new IndexShare mechanism, which shares one indexer across every four sparse attention layers (2.9× per-token FLOPs reduction at 1M context).
• Speculative decoding: an improved MTP layer increases acceptance length by up to 20%.
• Published model configuration: 78 hidden layers (the first 3 dense, the remainder sparse MoE), hidden size 6144, 256 routed experts with 8 routed experts active per token plus 1 shared expert, MoE intermediate size 2048, dense-layer intermediate size 12288, 64 attention heads, MLA-style attention with Q LoRA rank 2048 and KV LoRA rank 512, vocabulary size 154,880, and a maximum position embedding of 1,048,576 tokens.

Thinking and Reasoning Effort

Z.ai's documentation describes GLM-5.2's reasoning controls as follows:

• Deep thinking is enabled by default. GLM-5.2 uses dynamic thinking: the model automatically determines whether to think. Thinking can be disabled entirely for direct answers.
• Reasoning effort ladder. A reasoning_effort control (supported from GLM-5.2 onward) accepts max, xhigh, high, medium, low, minimal, and none. max is the default and recommended deep-reasoning level; high is enhanced reasoning; none or minimal skips thinking; for compatibility, low and medium are mapped to high, and xhigh is mapped to max.
• Generation defaults. Z.ai's API defaults for GLM-5.2 are temperature 1.0 and top_p 0.95, with up to 128K maximum output tokens. Z.ai's own evaluations use temperature=1.0, top_p=0.95, and a maximum generation length of 163,840 tokens for reasoning tasks.
• Tool use and caching. Z.ai documents function/tool calling for the GLM-5 series, and context caching (automatic reuse of repeated prefixes at a reduced rate) across the GLM-5 series.

Long-Horizon Performance (from the Z.ai announcement)

• On FrontierSWE, which measures whether an agent can complete open-ended technical projects at the scale of hours to tens of hours (systems optimization, large-scale code construction, applied ML research), GLM-5.2 trails Claude Opus 4.8 by only 1%, while edging out GPT-5.5 by 1% and Claude Opus 4.7 by 11%.
• On PostTrainBench, where each agent is given an H100 GPU and evaluated by how much it can improve small models through post-training, GLM-5.2 outperforms both Claude Opus 4.7 and GPT-5.5, ranking second only to Claude Opus 4.8.
• On SWE-Marathon, an ultra-long-horizon software engineering benchmark (building compilers, optimizing kernels, developing production-grade services), GLM-5.2 trails Claude Opus 4.8 by 13% while remaining second only to the Opus series.
• On standard coding benchmarks, GLM-5.2 is the strongest open-source model, improving on GLM-5.1 by a wide margin: 81.0 vs. 63.5 on Terminal-Bench 2.1 and 62.1 vs. 58.4 on SWE-bench Pro. On Terminal-Bench 2.1 (81.0) it lands within a few points of Claude Opus 4.8 (85.0) while staying ahead of Gemini 3.1 Pro.

Benchmarks

Full benchmark table as published in the official GLM-5.2 model card. Scores marked with * are from the full set (the default report is the text-only subset).

Reasoning

Benchmark	GLM-5.2	GLM-5.1	Qwen3.7-Max	MiniMax M3	DeepSeek-V4-Pro	Claude Opus 4.8	GPT-5.5	Gemini 3.1 Pro
HLE	40.5	31	41.4	37	37.7	49.8*	41.4*	45
HLE (w/ Tools)	54.7	52.3	53.5	-	48.2	57.9*	52.2*	51.4*
CritPt	20.9	4.6	13.4	3.7	12.9	20.9	27.1	17.7
AIME 2026	99.2	95.3	97	-	94.6	95.7	98.3	98.2
HMMT Nov. 2025	94.4	94	95	84.4	94.4	96.5	96.5	94.8
HMMT Feb. 2026	92.5	82.6	97.1	84.4	95.2	96.7	96.7	87.3
IMOAnswerBench	91.0	83.8	90	-	89.8	83.5	-	81
GPQA-Diamond	91.2	86.2	90	93	90.1	93.6	93.6	94.3

Coding

Benchmark	GLM-5.2	GLM-5.1	Qwen3.7-Max	MiniMax M3	DeepSeek-V4-Pro	Claude Opus 4.8	GPT-5.5	Gemini 3.1 Pro
SWE-bench Pro	62.1	58.4	60.6	59	55.4	69.2	58.6	54.2
NL2Repo	48.9	42.7	47.2	42.1	35.5	69.7	50.7	33.4
DeepSWE	46.2	18	18	20	8	58	70	10
ProgramBench	63.7	50.9	-	-	47.8	71.9	70.8	39.5
Terminal Bench 2.1 (Terminus-2)	81.0	63.5	75	65	64	85	84	74
Terminal Bench 2.1 (Best Reported Harness)	82.7	69	-	-	-	78.9	83.4	70.7
FrontierSWE (Dominance)	74.4	30.5	-	-	29.0	75.1	72.6	39.6
PostTrainBench	34.3	20.1	-	-	-	37.2	28.4	21.6
SWE-Marathon	13.0	1.0	-	-	-	26.0	12.0	4.0

Agentic

Benchmark	GLM-5.2	GLM-5.1	Qwen3.7-Max	MiniMax M3	DeepSeek-V4-Pro	Claude Opus 4.8	GPT-5.5	Gemini 3.1 Pro
MCP-Atlas (Public Set)	76.8	71.8	76.4	74.2	73.6	77.8	75.3	69.2
Tool-Decathlon	48.2	40.7	-	-	52.8	59.9	55.6	48.8

Evaluation Notes (as published by Z.ai)

• Humanity's Last Exam (HLE) and other reasoning tasks: sampling parameters temperature=1.0, top_p=0.95; maximum generation length 163,840 tokens. By default the text-only subset is reported; results marked with * are from the full set. For AIME, HMMT, and IMOAnswerBench, each question is evaluated with a system prompt asking for an explanation, an exact answer, and a confidence score between 0% and 100%; GPT-5.5 (medium) is used as the judge model. For HLE-with-tools, a maximum context length of 300,000 tokens is used with no context management strategy.
• SWE-bench Pro: run with the OpenHands agent framework using a tailored instruction prompt; temperature=1, top_p=1, max_new_tokens=32k, with a 400K context window.
• NL2Repo: evaluated with temperature=1.0, top_p=1.0, max_new_tokens=48k under 400k context. To prevent hacking, rule-based and LLM-based judgement is used to prevent malicious behaviors (e.g., unauthorized package-installation or download commands).
• DeepSWE: run with the official pier evaluation framework and the mini-swe-agent harness (temperature=1.0, top_p=1.0, timeout=2h, 400K context). Each task is solved in an isolated container with 2 CPUs, 8 GB RAM, and no internet access.
• ProgramBench: 200 instances evaluated with the Claude Code agent (version 2.1.156) using temperature=1.0, top_p=1.0, max_tokens=64000, max_turns=2000, sample_timeout=6h, reasoning_effort=max, with a 400K context window. Each instance runs in a (4 CPUs, 8 GB RAM) sandbox with internet access disabled.
• Terminal-Bench 2.1 (Terminus 2): evaluated with the Terminus-2 framework using parser=json, timeout=4h, temperature=1.0, top_p=1.0, max_new_tokens=48k, max_episodes=500, with a 256K context window. Resource limits are capped at 4 CPUs and 8 GB RAM.
• Terminal-Bench 2.1 (Claude Code harness): evaluated in Claude Code 2.1.167 with temperature=1.0, top_p=0.95, max_new_tokens=131072; the 64k CLI output cap is overridden to 128k via a transparent proxy. Wall-clock time limits are removed while per-task CPU and memory constraints are preserved. Scores are averaged over 5 runs.
• MCP-Atlas: all models evaluated in think mode on the 500-task public subset with a 10-minute timeout per task; Gemini-3.0-Pro is the judge model.
• Tool-Decathlon: official evaluation service with max_token set to 128K.
• FrontierSWE: evaluation conducted by Proximal with 1M context length, max effort level, and 128K maximum output tokens; dominance score reported as of 2026/06/16.
• PostTrainBench: evaluation conducted by the PostTrainBench project with 1M context length, max effort level, and 128K maximum output tokens.
• SWE-Marathon: evaluation conducted by Abundant AI with 1M context length, max effort level, and 128K maximum output tokens.

Training Notes

GLM-5.2 builds on the GLM-5 series. Per Z.ai's GLM-5 series materials: the series scales from GLM-4.5's 355B parameters (32B active) to 744B parameters (40B active) and increases pre-training data from 23T to 28.5T tokens; it integrates DeepSeek Sparse Attention (DSA) to largely reduce deployment cost while preserving long-context capacity; and post-training uses "slime", Z.ai's asynchronous reinforcement-learning infrastructure, which substantially improves training throughput and efficiency and enables more fine-grained post-training iterations. GLM-5.2 additionally proposes IndexShare (described in a separate Z.ai preprint) and an improved MTP layer. The GLM-5 technical report, "GLM-5: from Vibe Coding to Agentic Engineering," covers the series in detail.

License

GLM-5.2 is released under the MIT license — described in the announcement as "Pure Open: no regional limits, technical access without borders."

On Run BiOS

On the Run BiOS serverless catalog this model is served as glm-5.2 with its full 1,048,576-token context window, function calling, adjustable reasoning effort (exposed as a low / medium / high ladder; per Z.ai's mapping, low and medium run at the high level upstream, and the model's own default is the maximum deep-reasoning level), and prompt caching for repeated prefixes.