GLM-5.2: Open Long-Horizon Agents at 1M Context

Z.ai released GLM-5.2 on June 16, 2026 as an MIT-licensed, one-million-context model focused on long-horizon agents. The official release reports 81.0 on Terminal-Bench 2.1 with Terminus-2, 62.1 on SWE-bench Pro, 76.8 on MCP Atlas, 34.3 on PostTrainBench, and 13.0 on SWE-Marathon. It improves substantially over GLM-5.1 while remaining below the strongest closed models on several extended tasks.

The architectural headline is IndexShare. A shared indexer serves four sparse-attention layers, reducing redundant selection work. Z.ai reports a 2.9x reduction in per-token FLOPs at one million tokens, plus a 20 percent increase in multi-token-prediction acceptance. The model also exposes effort levels so applications can trade latency and token use for deeper work.

The release is aimed at completion, not chat

GLM-5.2's benchmark selection tells the product story. Terminal-Bench, SWE-bench Pro, MCP Atlas, PostTrainBench, and SWE-Marathon all require action over time. The model is designed to maintain plans, operate tools, and produce artifacts rather than answer isolated questions.

The weights are available under MIT, making the checkpoint more permissive than many “open” releases. Hosted access adds managed effort settings and tools. Raw weights, Z.ai API, and a third-party coding agent are distinct systems.

One million tokens support repositories and long traces. The release tries to make that window usable through sparse indexing rather than treating context length as a specification alone.

Full benchmark snapshot

The first-party table is unusually detailed:

Competitor columns in the same table mix first-party, API, and published results. The cleanest evidence is the matched GLM-5.1 comparison. Z.ai also distinguishes Terminus-2 from “best reported harness,” a valuable methodology practice.

IndexShare attacks repeated sparse-selection cost

Sparse attention first decides which blocks of a long history deserve expensive attention. If every layer builds a new index, selection itself becomes costly. IndexShare reuses one indexer across four sparse layers, amortizing the work.

Z.ai reports 2.9x fewer per-token FLOPs at one million context. FLOPs are not wall-clock latency. Index building, memory access, communication, kernels, and batch shape still matter. The shared index may also constrain each layer's ability to choose a different view of context; model training must compensate.

Test prefill, decode, and retrieval quality separately. Place related evidence across distant regions and introduce updates. A fast shared index is valuable only if it continues to select the right blocks.

Terminal and repository gains

The jump from 63.5 to 81.0 on Terminal-Bench 2.1 is the clearest same-family improvement. Terminus-2 provides the scaffold in this row. A separate best-reported-harness row gives different values, proving that the agent wrapper changes the outcome.

SWE-bench Pro improves more modestly to 62.1. ProgramBench rises strongly to 63.7. These results suggest better environment use and behavioral reasoning, but not automatic patch quality. Local tests should grade minimal diffs, tests, security, and reviewer effort.

The software-engineering agent guide recommends sandboxed tools, repository boundaries, and objective verification. Those controls apply even when a model posts a high terminal score.

Long-horizon tests reveal remaining distance

PostTrainBench gives an agent an H100 and asks it to improve small models. SWE-Marathon includes compiler, kernel, and production-service projects. GLM-5.2's gains over 5.1 are large, but the absolute Marathon score of 13 remains below Opus 4.8's 26 in Z.ai's table.

This is healthy evidence. The release does not hide where the open model trails. Long tasks magnify planning drift, bad checkpoints, and inefficient exploration. A model can excel on a two-hour terminal benchmark and still struggle over days.

Production pilots should include interrupted runs, bounded budgets, resumability, and intermediate artifacts. Score partial value and recovery, not only final pass. A failed 20-hour run is operationally different from a failed five-minute task.

Effort control and token economics

GLM-5.2 exposes effort levels, including Max, so callers can allocate more compute to difficult tasks. Z.ai's curves show better coding performance at comparable token budgets than GLM-5.1 and place higher effort between Opus generations under the reported setup.

Effort is a routing tool. Routine classification should not pay for Max; a hard repository migration might. Evaluate each tier at equal accepted quality and record output, hidden reasoning if billed, tool calls, and wall time.

Multi-token-prediction acceptance reportedly rises 20 percent, which can improve decode throughput. Acceptance depends on workload and drafter. Measure it in the chosen serving engine rather than assuming the vendor percentage.

Open deployment and best fit

MIT weights enable private serving, fine-tuning, and inspection. The model's scale and million-token context still demand significant infrastructure. Pin revision, tokenizer, chat template, sparse-attention kernels, and precision. Quantization needs its own multilingual and tool regression.

The official GLM-5.2 model card is the artifact record for that work. Confirm that the hosted model and downloaded revision use comparable prompts before treating API and self-hosted results as interchangeable. Even a tokenizer or effort-default change can alter cost and benchmark behavior.

GLM-5.2 is a strong candidate for open coding agents, long repository analysis, MCP automation, and research on long-horizon post-training. It is less suitable for small-device use or teams unable to run a large checkpoint. Hosted access may be more economical at low volume.

Use the open-model operations guide to control artifacts and rollout, and the frontier evaluation guide to keep harness versions visible.

Acceptance should include sparse-retrieval adversarial cases: duplicate file names, stale summaries, a late correction, and instructions embedded in retrieved documents. Record which context spans the model cites and whether IndexShare repeatedly misses a minority detail. For agent work, simulate rate limits, failed tools, and resume after interruption. These tests connect the architecture claim to production behavior.

Verdict

GLM-5.2 is one of the strongest open-weight releases in the window. It pairs a concrete long-context systems idea with large matched gains in terminal, program, and long-running work. Its detailed table makes both progress and remaining gaps visible.

The defensible conclusion is not that 81.0 proves universal coding leadership. It is that GLM-5.2 materially advances the GLM family and offers a permissive, configurable base for serious agents. Reproduce Terminal-Bench-like work, million-token retrieval, and day-scale tasks under the intended hardware before production.

Publish the local harness manifest with the result so later upgrades remain comparable.

Source notes — reviewed 2026

• Z.ai's GLM-5.2 release supplies the date, architecture, complete benchmark table, and effort controls.
• The official GLM-5.2 model card is the first-party weight, license, and configuration source.
• The official GLM repository provides deployment code and model lineage.
• Z.ai API documentation documents hosted access and effort settings.
• Terminal-Bench's official site documents the versioned terminal benchmark cited by Z.ai.