undefined

Posts

undefined

Followers

User since undefined NaN
Technology

Z.ai Releases GLM-5.1: 754B Model Tops SWE-Bench Pro

By Markus Kasanmascheff,

2026-04-09
https://img.particlenews.com/image.php?url=0ejSxW_1AJaVFrt00TL;DR
  • Release: Chinese AI lab Z.ai released GLM-5.1, a 754B-parameter open-weights model under MIT license.
  • Benchmark Lead: GLM-5.1 tops SWE-Bench Pro at 58.4%, outpacing Claude Opus 4.6 and GPT-5.4 on coding benchmarks.
  • Autonomous Execution: The model sustains independent task execution for over 8 hours across hundreds of rounds and thousands of tool calls.
  • API Access: Eleven providers on OpenRouter offer GLM-5.1 API access with a 202,752 token context window.
  • Open Strategy: Z.ai continues publishing weights under permissive licenses while Meta reportedly pivots away from open-source models.

Chinese AI lab Z.ai released GLM-5.1 on April 7. A 754B-parameter open-weights model, GLM-5.1 achieves top performance on SWE-Bench Pro and sustains autonomous task execution for over 8 hours.

Availability extends to Hugging Face under an MIT license and through 11 providers on OpenRouter.

GLM-5.1 marks the fourth major release from Z.ai in under twelve months. Pairing large-scale architecture with unrestricted access for developers represents a notable step in long-horizon agentic tasks.

A rapid succession of launches precedes this release: Zhipu AI rebranded as Z.ai in July 2025, launched GLM-4.5, followed by GLM-4.7 in December 2025, GLM-5 in February 2026, and now GLM-5.1. Just five days earlier on April 2, the company launched GLM-5V-Turbo, a multimodal vision model.

Benchmark Performance and Autonomous Capabilities

GLM-5.1 reaches 58.4% on SWE-Bench Pro. That score leads all current models, outpacing Claude Opus 4.6 at 57.3% and GPT-5.4 at 57.7%.

Qwen3.6-Plus lands at 56.6%, with Minimax M2.7 at 56.2%. According to the Hugging Face model card, CyberGym and BrowseComp benchmarks show similar leadership at 68.7% and 68.0% respectively.

On BrowseComp with Context Management, GLM-5.1 scores 79.3%. It trails Opus 4.6 at 84.0% and Gemini 3.1 Pro at 85.9%.

Improvements from predecessor GLM-5 prove substantial across coding benchmarks. CyberGym jumped 20.4 points. Terminal-Bench 2.0 rose 7.3 points.

NL2Repo climbed 6.8 points, and SWE-Bench Pro gained 3.3 points. Z.ai describes GLM-5.1 as its flagship agentic engineering model with notably stronger coding capabilities than GLM-5, per the Hugging Face model card.

Endurance stands out as the defining capability. GLM-5.1 can work independently for more than 8 hours.

The model autonomously plans, executes, and improves itself across hundreds of rounds and thousands of tool calls. Breaking complex problems down, running experiments, reading results, and identifying blockers with real precision defines the approach.

Earlier models including GLM-5 tend to exhaust their repertoire early and plateau. GLM-5.1 sustains optimization over hundreds of rounds without degrading.

An 8-hour threshold transforms what autonomous agents can attempt: multi-file architectural changes, iterative debugging sessions, and complex deployment workflows that previously required constant human oversight.

Math and reasoning benchmarks tell a more nuanced story. GLM-5.1 reaches 95.3% on AIME 2026 and 86.2% on GPQA-Diamond.

According to Hugging Face benchmark data, HLE scores sit at 31.0%. NL2Repo results show the model at 42.7%, up from GLM-5’s 35.9%, though Claude Opus 4.6 leads at 49.8%.

On Vending Bench 2, an economic benchmark measuring autonomous earning potential, GLM-5.1 earns $5,634.41. Opus 4.6 reaches $8,017.59 and GPT-5.4 hits $6,144.18.

https://img.particlenews.com/image.php?url=3tmGaQ_1AJaVFrt00

GPT-5.4 leads on five benchmarks including AIME 2026, HMMT Feb 2026, and Tool-Decathlon, while Gemini 3.1 Pro tops HLE, GPQA-Diamond, and Terminal-Bench 2.0. The benchmark spread reveals a model optimized for coding and agent workloads rather than universal dominance.

Model Specifications and API Access

Behind those benchmark scores sits a substantial model. Weighing in at 1.51TB on Hugging Face, the 754B parameter GLM-5.1 ranks among the largest open-weights models available for download.

Open weights under an MIT license allow developers to use, modify, and distribute the model without restrictive usage terms. The approach continues the open strategy that Zhang Peng articulated in February 2026: “Unlike OpenAI’s closed system, we adopt an open strategy.”

Developers unable to run 1.51TB locally can access GLM-5.1 via API through OpenRouter. The platform offers a 202,752 token context window.

GLM-5.1 API pricing on OpenRouter is set at $1.26 per million input tokens and $3.96 per million output tokens. Eleven providers carry the model on the platform, including a standard variant hosted on GMICloud using FP8 quantization.

Configurable reasoning tokens let developers adjust thinking depth based on task complexity.

Architecture and Training Innovations

A research paper authored by 184+ researchers from the GLM-5-Team documents the architecture enabling these capabilities. Aohan Zeng serves as primary contact.

Titled GLM-5: from Vibe Coding to Agentic Engineering, the paper was submitted in February 2026 and covers the architecture shared by both GLM-5 and GLM-5.1. Code and models are available at github.com/zai-org/GLM-5.

Dynamic Sparse Attention technology reduces computational costs while preserving long-context capabilities. The approach avoids the quadratic compute costs that typically force early truncation in extended sessions.

Asynchronous reinforcement learning infrastructure decouples generation from training processes. Multi-step interactions drive learning rather than isolated prompt-response pairs.

Novel asynchronous agent RL algorithms enable learning from complex, long-horizon interactions. Core capabilities focus on three areas: agentic functionality, reasoning abilities, and coding proficiency.

Z.ai states in its model card that “The longer it runs, the better the result”, reflecting the philosophy behind the extended autonomous execution design.

Hands-On Testing: The Pelican Test

Beyond benchmarks, early hands-on testing reveals how GLM-5.1 performs in practice. Simon Willison, an independent tech commentator who runs a standardized pelican-on-a-bicycle SVG generation test across AI models, encountered something new with GLM-5.1.

Testing via OpenRouter’s API using the llm CLI tool with the llm-openrouter plugin, Willison asked the model to generate an SVG of a pelican on a bicycle. GLM-5.1 produced not just the image but an entire animated HTML page with CSS animations, unprompted.

Animation initially failed. The pelican floated to the top-left corner of the screen.

Willison issued a one-line follow-up prompt describing the glitch, and GLM-5.1 replied with a precise technical diagnosis and corrected HTML, fixing itself in a single turn.

“The issue is that CSS transform animations on SVG elements override the SVG transform attribute used for positioning, causing the pelican to lose its placement and fly off to the top-right. The fix is to separate positioning (SVG attribute) from animation (inner group) and use <animateTransform> for SVG rotations since it handles coordinate systems correctly.”

GLM-5.1, AI model (Simon Willison’s Weblog)

Willison called the static SVG output his new favorite from an open-weights model. He made an exception to his usual single-prompt rule to test the model’s self-correction ability.

Following the animation fix, a Bluesky user suggested testing a second scene, and GLM-5.1 generated a complex animated opossum on an e-scooter. The output included detailed SVG comments like “Distant treeline silhouette – Virginia pines” and “Front paw on handlebar.”

The self-correction loop demonstrated here is what makes long-horizon execution viable: a model that diagnoses its own errors and produces targeted fixes without starting over differs fundamentally from one requiring human intervention at each failure point.

Prior Coverage and Context

GLM-5.1 shares the same paper as GLM-5. The relationship suggests an incremental update or fine-tune rather than a completely new architecture.

GLM-5 was a 744B-parameter model with a Mixture of Experts design. It used only 40 billion active parameters per inference despite its substantial total parameter count. Training occurred entirely on Huawei Ascend chips using the MindSpore framework, achieving full independence from US-manufactured semiconductor hardware.

GLM-5.1’s slight increase to 754B parameters suggests architectural refinement rather than a fundamental redesign.

Z.ai raised prices for GLM-5.1, going against the broader Chinese AI market trend of aggressive price cuts. According to reports, Zhipu AI annual revenue more than doubled, though losses climbed 60% amid intensifying competition in the Chinese AI sector.

The pricing decision signals confidence in the model’s differentiation, though widening losses suggest the open-weights approach carries notable compute costs that have yet to be offset by licensing or API revenue.

Z.ai’s commitment to open weights stands in contrast to Meta’s pivot away from open-source Llama models toward proprietary alternatives. While Meta retreats from its open-source philosophy, Z.ai continues to publish weights under permissive licenses, positioning itself as a key player in the open-weights ecosystem.

What Comes Next

GLM-5.1 leads in coding and agentic tasks while trailing GPT-5.4 and Gemini 3.1 Pro on math and general reasoning benchmarks across the broader market. On tool-use benchmarks, GLM-5.1 scores 71.8% on MCP-Atlas Public Set, 40.7% on Tool-Decathlon, and 70.6% on tau3-Bench, showing solid but not class-leading performance.

The release adds another major player to the open-weights competition. Developers gain access to frontier-scale coding capabilities without the usage restrictions that accompany proprietary models.

For teams building autonomous coding agents, the combination of MIT licensing, extended task endurance, and API accessibility through multiple providers creates a practical path to testing long-horizon workflows without vendor lock-in.

The post Z.ai Releases GLM-5.1: 754B Model Tops SWE-Bench Pro appeared first on WinBuzzer.

Comments