Z.ai ✓

Native multimodal training enables text, image, video, file, chart, and interface understanding within the same agent workflow.
Stronger coding and tool use outperform GLM-5.2 across repository engineering, terminal work, automation, and multi-step agent benchmarks.
Hybrid sparse and linear attention preserves precise long-context performance while cutting attention compute by 3x and KV-cache size by 4.4x.
A 1M-token context window supports large repositories, extended agent trajectories, and long professional workflows with substantially lower serving overhead.
320B total parameters with only 18B active delivers higher capability at roughly one-tenth the task cost while improving inference efficiency.

Delivers about 50% higher coding performance on Z.ai Code Bench while using fewer output tokens than the previous version.

Stronger long-horizon execution handles realistic engineering and research tasks that can require hours or days of expert work.

Major gains in terminal, repository, and software-engineering benchmarks improve autonomous coding across complex production environments.

Expanded cybersecurity reasoning improves vulnerability discovery and multi-stage exploitation, more than doubling performance on deeper exploitation benchmarks.

Scaled post-training improves coding, tool use, and agentic performance without changing the underlying base model.

Expands usable context from 200K to a stable 1M tokens, enabling much longer coding, research, debugging, and engineering workflows.

Stronger coding substantially improves repository-level work, terminal tasks, program generation, and long-horizon software engineering.

Adjustable reasoning effort lets users trade speed and cost for higher performance on difficult coding and agent tasks.

More efficient long-context architecture cuts per-token computation by 2.9x at 1M context while improving speculative decoding acceptance by up to 20%.

Improved agentic training strengthens multi-step tool use, task decomposition, long trajectories, and resistance to reward-hacking shortcuts.

Sustains autonomous work on a single task for up to 8 hours, covering planning, execution, testing, fixing, and final delivery.

Stronger agentic coding improves long-horizon planning, stepwise execution, strategy adjustment, and complex multi-stage engineering work.

Autonomous optimization loops can run benchmarks, identify bottlenecks, revise strategies, and repeatedly improve systems without manual intervention.

Better general assistance strengthens complex instruction following, long-context understanding, multi-turn reliability, and open-ended question answering.

Higher-quality production work improves websites, front-end prototypes, documents, presentations, PDFs, and spreadsheets with more polished default outputs.

GLM-5-Turbo is a lightweight, low-latency variant of the GLM-5 family designed for high-throughput chat and application workloads.

Optimized for cost-efficient inference, enabling large-scale deployments like assistants, agents, and real-time apps.

Supports long-context conversations and structured tool use, making it suitable for agent pipelines and coding workflows.

Provides strong multilingual performance, particularly for Chinese and English usage scenarios.

Available through the Z.ai API ecosystem, targeting developers building production LLM applications.

Repositioned as the new flagship for “agentic engineering,” targeting longer-horizon agent tasks end to end.

Introduces a dedicated programming-enhanced variant (“GLM-5-Code”) for more stable tool-invoking coding agents.

Stronger front-end generation accuracy called out explicitly for pages, animations, minigames, and 3D outputs.

Coding and agent performance is claimed as open-source SOTA, with usability approaching Claude Opus 4.5.

Keeps the same 200K context baseline while shifting the focus to higher reliability on long-range execution.

More stable multi-step reasoning and execution, tuned for “task completion” workflows.

Bigger jump in agentic coding and terminal-based tasks (reported benchmark gains on SWE-bench and Terminal Bench).

Stronger tool invocation, with higher reported scores on web browsing and interactive tool benchmarks.

Better front-end aesthetics and UI generation quality by default (webpages, slides, posters).

Added improvements to reasoning-control methods (retained and round-based reasoning) for more controllable complex runs.

Context window increased from 128K to 200K for heavier agent tasks.

Over 30% better token efficiency in real-world coding tests.

Stronger coding performance, including better results in common coding-agent environments.

Better reasoning plus explicit tool use during inference for stronger overall capability.

More capable tool and search agents, with tighter integration into agent frameworks.

Hybrid reasoning with “thinking” and “non-thinking” modes for complex reasoning and quick replies.

Built to unify reasoning, coding, and agent workflows in one model family.

Native function calling aimed at tool-using agents.

Long context support (128K) for larger prompts and agent traces.

Open weights released under an MIT-style license for commercial use.