GLM-5.1

GLM-5.1 is Z.AI’s next-generation open-weight foundation model, released on March 27, 2026 under the MIT License. It is the successor to GLM-5 (“From Vibe Coding to Agentic Engineering”) and sets a new state-of-the-art result on SWE-Bench Pro — outperforming GPT-5.4, Claude Opus 4.6, and Gemini 3.1 Pro on real-world software engineering tasks. Its defining characteristic is sustained long-horizon agentic execution: GLM-5.1 can work autonomously on a single complex task for up to 8 hours, continuously evaluating its own intermediate results and revising its approach hundreds of times before delivering final output.

Architecture

GLM-5.1 is a Sparse Mixture-of-Experts model with 754 billion total parameters and 40 billion active per token. The MoE layer uses 256 routed experts with top-8 routing plus 1 always-active shared expert, giving the model a stable base representation alongside dynamic specialisation.

Multi-head Latent Attention (MLA) GLM-5.1 replaces standard multi-head attention with MLA, which compresses key-value representations into a lower-dimensional latent space before projecting back out for attention computation. This reduces KV cache memory by roughly 33% versus standard MHA. The MLA-256 variant used here increases head dimension from 192 to 256 while reducing total attention heads by a third — cutting decoding compute without sacrificing training efficiency or context fidelity.

DeepSeek Sparse Attention (DSA) For long-context processing, GLM-5.1 combines MLA with DeepSeek Sparse Attention, which attends selectively across the 200K-token context window rather than computing full quadratic attention across all positions. This combination keeps long-context inference compute tractable at production scale.

Multi-Token Prediction (MTP) head A secondary prediction head is trained to predict multiple future tokens per forward pass. During inference, this enables speculative decoding — generating candidate continuations in parallel and verifying them in bulk — significantly improving tokens-per-second throughput without affecting output quality.

Training: Slime Asynchronous RL

GLM-5.1’s post-training uses Slime, Z.AI’s asynchronous reinforcement learning infrastructure. Standard RL for LLMs is sequential: generate a rollout, evaluate it, update weights, repeat — with the cluster idling while evaluation completes. Slime breaks the dependency:

• Training trajectories are generated independently across the cluster in parallel
• APRIL (Active Partial Rollouts) allows partial evaluation of incomplete trajectories, feeding results back into weight updates without waiting for full rollouts to finish
• This keeps GPU utilisation high throughout training and dramatically increases the volume of RL signal the model receives per wall-clock hour

The RL training distribution for GLM-5.1 was heavily weighted toward coding tasks. Crucially, long-horizon agentic data was incorporated during mid-training rather than only at the fine-tuning stage — which is the key reason the model generalises to multi-step agent workflows rather than excelling only at single-shot code generation.

Thinking mode is enabled by default in GLM-5.1, a change from GLM-5 where it was opt-in.

Agentic Execution

GLM-5.1 was explicitly designed for tools like Claude Code and OpenClaw. Its long-horizon planning improves on GLM-5 across the full agentic loop:

1. Planning — decompose a complex task into an ordered sequence of sub-tasks
2. Stepwise execution — invoke tools, write code, run tests, read outputs
3. Process adjustment — evaluate interim results and revise the plan mid-execution (potentially hundreds of times on a single task)
4. Result delivery — produce production-ready output after full self-verification

The model sustains this loop for up to 8 hours — a threshold that covers most real-world engineering tasks from specification to working, tested code.

Benchmarks & Pricing

API pricing: approximately $3 per million output tokens — roughly 94.6% of Claude Opus 4.6’s performance at a fraction of the cost.

Features supported: function calling, structured output, context caching, MCP integration, and native thinking mode.

Significance

GLM-5.1 is part of a broader pattern in 2026 where open-weight models reach or exceed proprietary frontier performance on specific capability domains — in this case, agentic software engineering. Its combination of MLA efficiency, sparse attention, asynchronous RL training, and native agentic data during mid-training represents a coherent architectural bet that the bottleneck for long-horizon agents is training distribution, not raw parameter count.