GLM-5.1 released: 754B agentic model outperforms Claude on coding benchmarks
Zhipu AI released GLM-5.1, a 754-parameter model optimized for agentic engineering tasks. The model scores 58.4% on SWE-Bench Pro, outperforming Claude 3.5 Sonnet (57.3%), and demonstrates sustained reasoning capability over hundreds of iterations.
GLM-5.1: 754B Agentic Model Outperforms Claude on Coding Benchmarks
Zhipu AI released GLM-5.1, a 754-parameter flagship model built for agentic engineering tasks. The model achieves 58.4% on SWE-Bench Pro—the primary metric for software engineering capability—exceeding Claude 3.5 Sonnet (57.3%) and maintaining substantial leads on specialized benchmarks.
Key Performance Metrics
GLM-5.1 achieves state-of-the-art performance across multiple agentic benchmarks:
- SWE-Bench Pro: 58.4% (vs. Claude 57.3%, Gemini 3.1 Pro 54.2%)
- NL2Repo (repo generation): 42.7% (vs. Claude 49.8%, significant improvement over GLM-5's 35.9%)
- Terminal-Bench 2.0: 63.5% on Terminus-2 suite
- CyberGym: 68.7% (vs. Claude 66.6%)
- BrowseComp with context management: 79.3% (vs. Gemini 84.0%, Claude 75.9%)
Mathematical reasoning shows mixed performance: 95.3% on AIME 2026 and 86.2% on GPQA-Diamond, trailing GPT-5.4 (98.7% on AIME) and Gemini 3.1 Pro (94.3% on GPQA).
Distinctive Agentic Capability
Unlike previous models including GLM-5, which plateau after initial optimizations, GLM-5.1 is designed to sustain effectiveness over extended problem-solving horizons. According to the developers, the model handles ambiguous problems with improved judgment and maintains productivity across longer sessions—breaking complex tasks into experiments, reading results, identifying blockers, and revising strategies through hundreds of iterations and thousands of tool calls.
This iterative reasoning approach distinguishes it from models optimized for single-pass performance.
Deployment and Quantization
Unsloth released GGUF quantized versions with 17 quant options ranging from 206 GB (1-bit UD-IQ1_M) to 1.51 TB (16-bit BF16). The releases implement Unsloth Dynamic 2.0 quantization, which the developers claim achieves superior accuracy compared to other quantization methods.
Supported inference frameworks include:
- SGLang (v0.5.10+)
- vLLM (v0.19.0+)
- xLLM (v0.8.0+)
- Transformers (v4.5.3+)
- KTransformers (v0.5.3+)
The model received 13,329 downloads on Hugging Face in its first month.
Availability
GLM-5.1 is available through Z.ai API Platform for inference. The developers announced chat.z.ai access would come in subsequent days. A technical report and GitHub repository were published alongside the release.
What This Means
GLM-5.1 represents a shift in agentic model design: instead of pursuing raw benchmark scores on isolated tasks, the focus is extended-horizon reasoning and iterative refinement. Its SWE-Bench Pro lead over Claude positions it as the strongest open-access model for software engineering tasks, though Gemini 3.1 Pro and GPT-5.4 maintain mathematical reasoning advantages. The quantized GGUF versions enable local deployment at scale, with memory requirements scaling from 206 GB to 1.51 TB depending on precision needs.
Related Articles
Poolside releases Laguna XS 2.1: 33B parameter MoE coding model with 262K context window
Poolside has released Laguna XS 2.1, a 33B total parameter Mixture-of-Experts model with 3B activated parameters per token and a 262,144-token context window. The model achieves 70.9% on SWE-bench Verified and 63.1% on SWE-bench Multilingual, representing a 5.4% improvement over its predecessor on multilingual coding tasks.
DeepSeek Releases V4-Flash: 284B Parameter MoE Model with 1M Context Window at Q8 162GB
Unsloth has released optimized GGUF quantizations of DeepSeek-V4-Flash, a 284B parameter Mixture-of-Experts model that activates 13B parameters and supports 1 million token context windows. The Q8 quantization (UD-Q8_K_XL) runs at 162GB with claimed lossless precision, only 7GB larger than the Q4 variant.
xAI releases Grok 4.5 at $2/$6 per million tokens, claims Opus 4.7 performance at 60% lower cost
xAI has released Grok 4.5, pricing it at $2 per million input tokens and $6 per million output tokens — significantly undercutting Anthropic's Opus 4.7 ($5/$25 per million). Elon Musk claims the model delivers comparable performance to Opus 4.7 while being faster and more token-efficient.
Nex AGI releases Nex-N2-Mini: open-source agentic MoE model with 262K context window
Nex AGI has released Nex-N2-Mini, an open-source agentic mixture-of-experts model with a 262K-token context window. The model accepts text and image inputs and is priced at $0.025 per 1M input tokens and $0.10 per 1M output tokens.
Comments
Loading...