GLM-5 Open Source Model: MIT License, Top Coding Score 2026

On February 6, 2026, an anonymous model called "Pony Alpha" appeared on OpenRouter with no identifying information. Within hours the AI community had clocked its performance -- strikingly close to Claude Opus 4.5 on coding tasks. When prompted with "who are you?" the model answered: "I am GLM." Five days later, Z.ai confirmed: Pony Alpha was GLM-5, a new open-source model trained on 28.5 trillion tokens and released free under the MIT license. The company's shares on the Hong Kong Stock Exchange jumped 60% in three days. That debut is a useful frame for understanding what GLM-5 represents in 2026: a model competitive with the best closed models in the world, available to anyone with a Hugging Face account.

What is GLM-5 and who built it?

GLM-5 comes from Z.ai, a Beijing-based company formerly known as Zhipu AI. Founded in 2019 as a spinoff from Tsinghua University, Z.ai rebranded internationally in 2025 and went public on the Hong Kong Stock Exchange (SEHK: 2513) in January 2026. The company is one of China's so-called "AI tiger" companies and the third-largest LLM market player in China according to the International Data Corporation.

The GLM name stands for General Language Model, a training algorithm Z.ai researchers published at ACL in 2022 using an "autoregressive blank infilling" strategy. GLM-5 is the fifth major iteration. It uses a Mixture-of-Experts (MoE) architecture with 744 billion total parameters and 40 billion active parameters per forward pass -- meaning that despite the enormous total parameter count, the actual compute cost per token is far lower than a dense model of the same size. The model was pre-trained on 28.5 trillion tokens, up from 23 trillion in its predecessor GLM-4.5.

One technical detail that carries significant geopolitical weight: GLM-5 was trained entirely on Huawei Ascend chips, with no NVIDIA hardware involved. The US Commerce Department had placed Z.ai (then Zhipu AI) on its Entity List in January 2025 due to national security concerns, restricting US companies from exporting certain technologies to the firm. GLM-5's Huawei-only training pipeline is a direct demonstration that China can develop frontier-class models without access to US-controlled hardware.

What benchmark scores has GLM-5 achieved?

GLM-5 and its successor GLM-5.1 post results that match or exceed several closed frontier models on key benchmarks. Here is how GLM-5 compares against leading models as of mid-2026, based on Z.ai's published benchmark data:

The standout number is the HLE (Humanity's Last Exam) with Tools score: at 50.4%, GLM-5 outperforms both Claude Opus 4.5 and GPT-5.2. This benchmark tests complex multi-step reasoning using external tool calls -- the kind of task that matters for real-world agentic workflows. For teams building autonomous agents, GLM-5's lead here is meaningful.

On the Code Arena leaderboard, GLM-5.1 holds an Elo rating of 1530, making it the top-ranked open-source model for coding tasks. The overall BenchLM.ai provisional score for GLM-5.1 is 82/100, placing it 19th out of 119 evaluated models -- ahead of many paid-only alternatives that cost $20 to $200 per month to access.

How can you download and run GLM-5 for free?

There are three practical ways to access GLM-5 without spending money:

• chat.z.ai (free web chat) -- Z.ai's hosted interface at chat.z.ai gives free access to GLM-5 with rate limits. No download required. Good for testing and casual use.
• Hugging Face weights (full precision) -- The full GLM-5 and GLM-5.1 weights are available at zai-org/GLM-5 on Hugging Face under MIT license. At 744B parameters, you need multi-GPU server hardware to run these at full precision. Intended for research labs, cloud deployments, and enterprises with GPU clusters.
• Unsloth GGUF quantized builds -- Unsloth maintains quantized GGUF versions that run via llama.cpp on consumer hardware. The IQ2_M variant is the smallest, and Unsloth's documentation provides step-by-step instructions for running it locally. This is the fastest path to local self-hosting.

For teams that need cloud deployment without managing their own GPU cluster, Spheron and other GPU-cloud providers offer one-click GLM-5.1 deployments billed by compute hour. NVIDIA has also released an optimized NVFP4 quantized version (nvidia/GLM-5.1-NVFP4 on Hugging Face) designed for high-throughput inference on NVIDIA hardware -- a somewhat ironic development given that GLM-5 itself was trained entirely on Huawei chips.

See our open-source AI models directory for a full list of MIT-licensed models you can self-host today.

How does GLM-5 compare to other leading open-source models?

The open-source AI landscape in mid-2026 is remarkably competitive. GLM-5 competes directly with several other major releases from the same period, all freely available:

• DeepSeek V4 -- Released April 24, 2026 under MIT license. 1.6T total / 49B active parameters. Strong on coding but GLM-5's HLE tool score edges it out on agentic tasks.
• MiniMax M2.5 -- Achieves 80.2% on SWE-bench Verified, matching Claude Opus 4.5. GLM-5.1 trails on raw SWE-bench but leads on Code Arena Elo.
• Kimi K2.5 -- SWE-bench at 76.8%, HLE with Tools at 51.8% (narrowly ahead of GLM-5). Kimi K2.5 is the closest peer on the overall agentic benchmark suite.
• Qwen 3.5 -- Apache 2.0 licensed, strong multilingual and coding capabilities, competitive at smaller parameter counts. Useful when deployment cost is the primary constraint.

According to the iternal.ai LLM selection guide, the gap between open-source and proprietary models has "effectively closed for coding tasks" in 2026. GLM-5.1's Code Arena Elo of 1530 sits alongside MiniMax M2.5's SWE-bench score as evidence of this shift. For developers who previously assumed they needed a Claude or GPT subscription for serious coding work, the free alternatives now rival or match those offerings on the metrics that matter most.

Check out our free vs paid AI model comparison to see exactly where each open-source model lands against its paid competitors.

What is GLM-5.1 and what changed from the original GLM-5?

GLM-5.1 is Z.ai's incremental update to GLM-5, released in the weeks following the original February 2026 launch. The architecture is the same -- 754B total parameters, 40B active -- but the post-training has been refined using Z.ai's proprietary slime RL framework, an asynchronous reinforcement learning infrastructure that the company developed to improve training throughput at scale.

The slime framework enables what Z.ai calls "more fine-grained post-training iterations" -- essentially more rounds of RLHF-style alignment and capability tuning than would be practical with traditional synchronous RL methods. The result is that GLM-5.1 shows incremental gains over GLM-5 on coding and agentic tasks without changing the underlying pre-trained model. Think of it as the same engine with better tuning.

GLM-5.1 also benefits from NVIDIA's NVFP4 quantization, which compresses the model's weights into a format optimized for NVIDIA GPU memory bandwidth. This makes full-scale cloud deployment significantly faster and cheaper on NVIDIA infrastructure -- again, somewhat ironic given GLM-5's chip-independent training story, but it demonstrates the model's pragmatic cross-platform reach.

What legal and geopolitical considerations should users know about?

GLM-5 is MIT licensed, which means nearly any use is permitted: commercial deployment, modification, redistribution, and self-hosting are all allowed with minimal restriction. For most individual developers and companies, there are no legal barriers to using GLM-5.

The more nuanced consideration involves Z.ai's place on the US Entity List. The Commerce Department blacklisted the company in January 2025, which restricts US persons from exporting certain controlled technologies to Z.ai. However, downloading and using the MIT-licensed model weights -- which Z.ai has publicly distributed -- is a different matter from exporting technology to Z.ai. Most legal analyses treat downloading and using open-source weights under a permissive license as distinct from the export control restrictions. That said, organizations in regulated sectors (defense, government contractors) should consult legal counsel before deployment.

There is also the "Pony Alpha" identity episode to consider. MIT researchers documented that GLM-series models identified themselves as Claude approximately 50% of the time when queried through non-standard methods -- a quirk that Z.ai has never publicly explained. It does not affect model performance, but it has been noted in enterprise evaluations where auditability of model identity matters.

For most developers, the practical reality is straightforward: GLM-5 and GLM-5.1 are freely available, MIT licensed, and benchmark-competitive with the world's best models. Track new open-source model releases as they happen via our free tier tracker.