GLM-5.1

Morphllm's technical deep-dive describes GLM-5.1 as Z.ai's open-weight (MIT license) mixture-of-experts coding model released April 7, 2026, with 744B total parameters and 40B active per token, a 200K-token context window, and 128K max output. It scored 58.4 on SWE-bench Pro — the top score at its April 2026 launch, ah The article details GLM-5.1's DSA sparse attention architecture, its 8-hour autonomous agentic training objective, and notes it runs on transformers, vLLM, SGLang, KTransformers, and xLLM, integrating with Claude Code or Cline. It also documents a practical "glm47 tool-call parser gotcha" and frames GLM-5.2 as a 1M-con

An independent model card for GLM-5.1 records developer-reported benchmark scores: SWE-bench Pro at 58.4% resolved, Terminal-Bench 2.0 at 63.5% tasks resolved, and GPQA Diamond at 86.2% accuracy. The card lists a 203K context window, a release date of 2026-04, and pricing reference points of $1 per million input tokens and $3.20 per million output tokens. The card also estimates the model runs on 754B parameters and maps hardware fit across consumer and data-center GPUs from 12GB RTX 3060 cards up through 96GB RTX PRO 6000 Blackwell and 128GB unified-memory M5 Max systems. Data is dated to 2026-04 and the page commits to fact-checked, dated pricing and benchmark provenance rather than promotional framing.

An independent technical analysis breaks down GLM-5.1 as an open-source flagship LLM built around a Mixture-of-Experts architecture that keeps parameter scale above 100B while activating roughly 40B per inference. The MoE design pairs 256 routing experts with 1 shared expert under a Top-8 activation strategy, using a Gumbel-Softmax-based gating network to balance utilization, which reportedly lifts code-generation expert usage to 92% versus traditional Top-2 routing. The analysis adds that GLM-5.1 ships under the MIT license, giving developers rights to modify, redistribute, and commercialize the weights. It frames GLM-5.1 as differentiated on parameter scale, architectural innovation, and permissive open-source strategy versus peer models, and highlights the model's positioning for complex multi-turn interactions and long-sequence reasoning. Three specialized attention mechanisms are also integrated to support long-context inference efficiency.

GLM-5.1 is Zhipu AI's (Z.ai) follow-up point release to GLM-5, with its Hugging Face repository created on 2026-04-03 and public release dated 2026-04-07. The model scales to 744B total parameters with 40B active per inference, up from GLM-5's 355B/32B, and is distributed under the MIT license on Hugging Face and ModelScope. The release targets agentic engineering and long-horizon tasks, sustaining optimization across hundreds of rounds and thousands of tool calls, and reaching state-of-the-art performance on SWE-Bench Pro while leading on NL2Repo and Terminal-Bench 2.0. It inherits GLM-5's DeepSeek Sparse Attention (DSA) for cost-efficient long-context inference and the asynchronous reinforcement-learning infrastructure called "slime," documented in arXiv 2602.15763. Both GLM-5 and GLM-5.1 serve via the Z.ai API with a 200K context window and 128K maximum output.

Z.ai's official release notes page (docs.z.ai) explicitly lists GLM-5.1 as a 2026-04-07 release, describing it as designed for long-horizon tasks where it can work independently for up to 8 hours in a single run, covering a full loop from planning and execution to iterative refinement and final delivery. The page state The same release-notes page situates GLM-5.1 within Z.ai's broader GLM lineage, showing it was followed by GLM-5.2 (2026-06-16, 1M context, open-source SOTA on coding and long-horizon benchmarks), GLM-5.3 (2026-08-18, 50% coding gain over 5.2, emergent cybersecurity capabilities), and GLM-5.3-Flash (2026-08-26, 320B/18