Z.AI, the AI platform developed by the crew behind the GLM mannequin household, has launched GLM-5.1 — its next-generation flagship mannequin developed particularly for agentic engineering. In contrast to fashions optimized for clear, single-turn benchmarks, GLM-5.1 is constructed for agentic duties, with considerably stronger coding capabilities than its predecessor, and achieves state-of-the-art efficiency on SWE-Bench Professional whereas main GLM-5 by a large margin on NL2Repo (repo technology) and Terminal-Bench 2.0 (real-world terminal duties).
Structure: DSA, MoE, and Asynchronous RL
Earlier than diving into what GLM-5.1 can do, it’s value understanding what it’s constructed on — as a result of the structure is meaningfully completely different from a typical dense transformer.
GLM-5 adopts DSA to considerably scale back coaching and inference prices whereas sustaining long-context constancy. The mannequin makes use of a glm_moe_dsa structure (Combination of Specialists (MoE) mannequin mixed with DSA). For AI devs evaluating whether or not to self-host, this issues: MoE fashions activate solely a subset of their parameters per ahead move, which might make inference considerably extra environment friendly than a comparably-sized dense mannequin, although they require particular serving infrastructure.
On the coaching aspect, GLM-5 implements a brand new asynchronous reinforcement studying infrastructure that drastically improves post-training effectivity by decoupling technology from coaching. Novel asynchronous agent RL algorithms additional enhance RL high quality, enabling the mannequin to be taught from advanced, long-horizon interactions extra successfully. That is what permits the mannequin to deal with agentic duties with the type of sustained judgment that single-turn RL coaching struggles to provide.
The Plateau Drawback GLM-5.1 is Fixing
To grasp what makes GLM-5.1 completely different at inference time, it helps to grasp a selected failure mode in LLMs used as brokers. Earlier fashions — together with GLM-5 — are inclined to exhaust their repertoire early: they apply acquainted strategies for fast preliminary positive factors, then plateau. Giving them extra time doesn’t assist.
It is a structural limitation for any developer making an attempt to make use of an LLM as a coding agent. The mannequin applies the identical playbook it is aware of, hits a wall, and stops making progress no matter how lengthy it runs. GLM-5.1, against this, is constructed to remain efficient on agentic duties over for much longer horizons. The mannequin handles ambiguous issues with higher judgment and stays productive over longer periods. It breaks advanced issues down, runs experiments, reads outcomes, and identifies blockers with actual precision. By revisiting its reasoning and revising its technique via repeated iteration, GLM-5.1 sustains optimization over a whole bunch of rounds and hundreds of device calls.
The sustained efficiency requires greater than a bigger context window. This functionality requires the mannequin to take care of aim alignment over prolonged execution, decreasing technique drift, error accumulation, and ineffective trial and error, enabling really autonomous execution for advanced engineering duties.
Benchmarks: The place GLM-5.1 Stands
On SWE-Bench Professional, GLM-5.1 achieves a rating of 58.4, outperforming GPT-5.4, Claude Opus 4.6, and Gemini 3.1 Professional, setting a brand new state-of-the-art consequence.
The broader benchmark profile reveals a well-rounded mannequin. GLM-5.1 scores 95.3 on AIME 2026, 94.0 on HMMT Nov. 2025, 82.6 on HMMT Feb. 2026, and 86.2 on GPQA-Diamond — a graduate-level science reasoning benchmark. On agentic and tool-use benchmarks, GLM-5.1 scores 68.7 on CyberGym (a considerable soar from GLM-5’s 48.3), 68.0 on BrowseComp, 70.6 on τ³-Bench, and 71.8 on MCP-Atlas (Public Set) — the final one significantly related given MCP’s rising position in manufacturing agent programs. On Terminal-Bench 2.0, the mannequin scores 63.5, rising to 66.5 when evaluated with Claude Code because the scaffolding.
Throughout 12 consultant benchmarks masking reasoning, coding, brokers, device use, and looking, GLM-5.1 demonstrates a broad and well-balanced functionality profile. This reveals that GLM-5.1 shouldn’t be a single-metric enchancment — it advances concurrently throughout normal intelligence, real-world coding, and complicated job execution.
By way of total positioning, GLM-5.1’s normal functionality and coding efficiency are total aligned with Claude Opus 4.6.
8-Hour Sustained Execution: What That Really Means
Crucial distinction in GLM-5.1 is its capability for long-horizon job execution. GLM-5.1 can work autonomously on a single job for as much as 8 hours, finishing the complete course of from planning and execution to testing, fixing, and supply.
For builders constructing autonomous brokers, this adjustments the scope of what’s doable. Moderately than orchestrating a mannequin over dozens of short-lived device calls, you’ll be able to hand GLM-5.1 a posh goal and let it run a whole ‘experiment–analyze–optimize’ loop autonomously.
The concrete engineering demonstrations make this tangible: GLM-5.1 can construct a whole Linux desktop atmosphere from scratch in 8 hours; carry out 178 rounds of autonomous iteration on a vector database job and enhance efficiency to 1.5× the preliminary model; and optimize a CUDA kernel, growing speedup from 2.6× to 35.7× via sustained tuning.
That CUDA kernel result’s notable for ML engineers: bettering a kernel from 2.6× to 35.7× speedup via autonomous iterative optimization is a stage of depth that will take a talented human engineer vital time to copy manually.
Mannequin Specs and Deployment
GLM-5.1 is a 754-billion-parameter MoE mannequin launched underneath the MIT license on HuggingFace. It operates with a 200K context window and helps as much as 128K most output tokens — each vital for long-horizon duties that want to carry massive codebases or prolonged reasoning chains in reminiscence.
GLM-5.1 helps pondering mode (providing a number of pondering modes for various eventualities), streaming output, perform calling, context caching, structured output, and MCP for integrating exterior instruments and knowledge sources.
For native deployment, the next open-source frameworks assist GLM-5.1: SGLang (v0.5.10+), vLLM (v0.19.0+), xLLM (v0.8.0+), Transformers (v0.5.3+), and KTransformers (v0.5.3+).
For API entry, the mannequin is on the market via Z.AI’s API platform. Getting began requires putting in zai-sdk through pip and initializing a ZaiClient together with your API key. .
Key Takeaways
- GLM-5.1 units a brand new state-of-the-art on SWE-Bench Professional with a rating of 58.4, outperforming GPT-5.4, Claude Opus 4.6, and Gemini 3.1 Professional — making it one of many the strongest publicly benchmarked mannequin for real-world software program engineering duties on the time of launch.
- The mannequin is constructed for long-horizon autonomous execution, able to engaged on a single advanced job for as much as 8 hours — working experiments, revising methods, and iterating throughout a whole bunch of rounds and hundreds of device calls with out human intervention.
- GLM-5.1 makes use of a MoE + DSA structure educated with asynchronous reinforcement studying, which reduces coaching and inference prices in comparison with dense transformers whereas sustaining long-context constancy — a significant consideration for groups evaluating self-hosting.
- It’s open-weight underneath the MIT license (754B parameters, 200K context window, 128K max output tokens) and helps native deployment through SGLang, vLLM, xLLM, Transformers, and KTransformers, in addition to API entry via the Z.AI platform with OpenAI SDK compatibility.
- GLM-5.1 goes past coding — it additionally reveals sturdy enhancements in front-end prototyping, artifacts technology, and workplace productiveness duties (Phrase, Excel, PowerPoint, PDF), positioning it as a general-purpose basis for each agentic programs and high-quality content material workflows.
Take a look at the Weights, API and Technical particulars. Additionally, be at liberty to comply with us on Twitter and don’t neglect to affix our 120k+ ML SubReddit and Subscribe to our E-newsletter. Wait! are you on telegram? now you’ll be able to be a part of us on telegram as nicely.
Must companion with us for selling your GitHub Repo OR Hugging Face Web page OR Product Launch OR Webinar and many others.? Join with us
