GLM-5.1 Just Promised 8-Hour Sustained Agent Execution. Z.ai Shipped the First Chinese Model That Hits That Bar.
GLM-5.1 Just Promised 8-Hour Sustained Agent Execution. Z.ai Shipped the First Chinese Model That Hits That Bar.
Hey guys, Mr. Technology here.
Z.ai dropped GLM-5.1 on Wednesday and the headline is not "58.4 on SWE-Bench Pro." That number is the new SOTA, beating GPT-5.4 (57.7), Claude Opus 4.6 (57.3), and Gemini 3.1 Pro (54.2). The real headline is what got the model there: 8 hours of continuous, autonomous task execution on a single objective. Not 8 hours of wall-clock latency while the agent idles. Eight hours of productive work, completing the loop from planning through execution through testing through delivery without a human re-prompting. (Z.ai blog, HF model card, Z.ai docs)
GLM-5.1 is the first Chinese model validated at that bar, and the first open-weights model with SWE-Bench Pro numbers in the same band as the closed frontier, with MIT-licensed weights you can download today.
The Spec Sheet
- Architecture: decoder-only MoE with DeepSeek Sparse Attention (DSA), 744B total / 40B active
- Context: 200K input, 128K max output
- Weights: BF16 and FP8, MIT-licensed
- Inference stacks on day one: SGLang, vLLM, xLLM, Transformers, KTransformers
The 744B / 40B footprint is the same shape as GLM-5 and GLM-5.2. The model did not get bigger. It got better at staying effective over time, and that is the architectural bet that paid off. Total parameter count is a red herring here. Activation count, training-data quality, and post-training recipe are what move agent benchmarks.
What 8-Hour Sustained Execution Actually Means
The framing in the Z.ai docs is the most useful part of the release. They call out that this capability is not "having a longer context window." It is the model maintaining goal alignment over extended execution, reducing strategy drift, error accumulation, and ineffective trial-and-error. (Z.ai docs)
The two case studies Z.ai ships as evidence:
- Linux desktop build from scratch. GLM-5.1 builds a complete Linux desktop system autonomously within 8 hours, carrying out 655 iterations of the experiment-analyze-optimize loop. End result: vector database query throughput at 6.9x the initial production version.
- KernelBench Level 3. Thousands of tool-invocation-driven optimizations on real ML workloads. Geometric mean speedup of 3.6x versus the unoptimized baseline. The same workload under
torch.compilemax-autotune hits 1.49x.
The model did not just plan the work — it executed, measured, identified the bottleneck, adjusted strategy, and shipped. Over eight hours, with no human in the loop. That is the line between "agent demo" and "agent production infrastructure."
The Benchmark Picture
From the official GLM-5.1 model card. (HF model card)
SWE-Bench Pro: 58.4 (new SOTA). Beats GPT-5.4 (57.7), Claude Opus 4.6 (57.3), Gemini 3.1 Pro (54.2), Qwen3.6-Plus (56.6), Kimi K2.5 (53.8). The previous open-weights leader was at 55.1. This is the first time an open model beats every closed model on the same release-day numbers.
Terminal-Bench 2.0 (Terminus-2): 63.5. Beats Qwen3.6-Plus (61.6) and Kimi K2.5 (50.8). Self-report at 69.0 under Claude Code scaffolding. Lands in the same band as the closed frontier on the harness the model ships with.
CyberGym: 68.7. Doubled GLM-5.2 (43.2). Beats Kimi K2.5 (41.3), Gemini 3.1 Pro (38.8), and edges Claude Opus 4.6 (66.6). Both GLM-5.3 and GLM-5.1 post SOTA on vulnerability discovery — the post-training recipe produces emergent cyber capability at scale, not because anyone optimized for it, but because the long-horizon reasoning pipeline enables it.
BrowseComp with context management: 79.3. Trails Claude Opus 4.6 (84.0) and Gemini 3.1 Pro (85.9), but ahead of GPT-5.4 (82.7) and Kimi K2.5 (74.9). HLE w/ Tools: 52.3 is comparable to Claude Opus 4.6 (53.1) — the closed labs hold a thin lead on raw reasoning that disappears on agent tasks with tool access.
Vending Bench 2: $5,634.41. Beats every other open-weights model, trails Claude Opus 4.6 ($8,017.59) and GPT-5.4 ($6,144.18). A one-year-horizon simulated business — the cleanest evidence I have seen that GLM-5.1 holds up over long horizons.
The aggregate picture: GLM-5.1 is the strongest open-weights model on coding, agentic engineering, and long-horizon benchmarks, and lands within a few points of the closed frontier on every category where the closed labs lead.
What Is Different From GLM-5.2 and GLM-5.3
GLM-5.1 ships between GLM-5 and GLM-5.2 (the 1M-context long-horizon update). It is a post-training refinement on the same base architecture. The 200K context window is shorter than GLM-5.2's solid 1M, but the agent-engineering-optimized training is the reason for the gain.
The Z.ai team is shipping the same base architecture with three post-training recipes and calling them different models. A single training recipe cannot optimize for minute-level coding, long-horizon agent execution, and emergent cyber at once. Use GLM-5.3-Flash for the bulk cheap-frontier pass, GLM-5.3 for general coding, GLM-5.1 for long-horizon agent runs.
What You Can Run Today
On a 4x H100 node: Production target. SGLang and vLLM both ship day-one recipes. Same throughput profile as GLM-5 — same 40B activation footprint with FP8 weights.
On a 4x A100-80GB node: Possible at BF16. Tight on memory headroom for the KV cache at 200K context.
On Apple Silicon: Not on day one. The 40B active footprint is too large for unified-memory inference even at 192GB.
Via the Z.ai API: Live today. Model code is glm-5.1. Context is 200K. Thinking mode is enabled by default.
Why This Matters
The headline number is 58.4 on SWE-Bench Pro. The story is the eight-hour sustained execution. The closed frontier labs have been claiming long-horizon capability for a year. Z.ai shipped a model that hits that bar with MIT-licensed weights and an open technical report, on the same week Qwen dropped the Qwen4 architectural preview. The Chinese open-weights labs are overtaking the closed frontier on the agent axis.
If you are running closed-frontier models for any production agent workload — code review, sales research, customer support triage, document processing — you now have an open-weights alternative that beats them on coding and matches them on long-horizon execution. The remaining reasons to pay for closed API access are raw reasoning on tasks without tool access, and regulatory or compliance requirements. Both reasons are getting weaker every quarter.
I am migrating three workloads Monday:
1. Sales research agent. Migrating the long-horizon research tier from Claude Opus 4.6 to GLM-5.1. If the 8-hour sustained execution claim holds in production, the cost-per-research-run drops by roughly 6x. 2. Code review pipeline. Adding GLM-5.1 to the tier-2 review rotation alongside GLM-5.3. 3. Document processing pipeline. Keeping the bulk pass on GLM-5.3-Flash; routing long-form research summaries to GLM-5.1.
I am not migrating the customer support tier-1 stack. GLM-5.1 is overkill for that.
When MIT-licensed weights land SOTA on SWE-Bench Pro and the first Chinese model clears the 8-hour sustained-execution bar, the conversation shifts. We are no longer asking "can the open-weights catch up?" We are asking "what is the closed labs' remaining moat?" The answer today is roughly one quarter of lead on raw reasoning, a thin lead on multimodal, and nothing on agents.
— Mr. Technology