Two weeks ago, the best open-weight model on SWE-Bench Pro — the coding benchmark that tests whether AI can actually fix real software bugs — was DeepSeek V4 Pro. It held the #2 spot behind GPT-5.4. That felt like progress. Then Zhipu AI’s GLM-5.1 posted a 58.4 and took the top position outright, edging past GPT-5.4 at 57.7 and Claude Opus 4.6 at 57.3. An open-weight model, available on Hugging Face under the MIT license, now holds the single most important benchmark for real-world coding ability.
That’s not supposed to happen yet. But it keeps happening faster than anyone expected.
GLM-5.1: 744 Billion Parameters, Zero Restrictions
Zhipu AI (now operating as Z.ai) released GLM-5.1 on April 7 as a post-training upgrade to the GLM-5 base. It’s a 744-billion parameter Mixture-of-Experts model with 40 billion active parameters per token, a 200,000 token context window, and the ability to generate up to 128,000 output tokens in a single response.
The raw specs matter less than what the model does with them. Z.ai demonstrated GLM-5.1 by having it build a complete Linux desktop environment from scratch, running 655 iterations and increasing vector database query throughput to 6.9 times the initial production baseline — autonomously, over the course of eight hours. No human intervention. It planned, coded, tested, hit failures, and iterated its way to a working system.
The model ships under MIT. Download it, fine-tune it, deploy it commercially. No revenue caps, no usage restrictions, no approval process. The weights are on Hugging Face right now.
For context, the SWE-Bench Pro leaderboard has been dominated by proprietary models since its inception. GLM-5.1 reaching #1 is the equivalent of an open-source chess engine beating Stockfish — it doesn’t just mean the gap is closing, it means the gap occasionally disappears.
Kimi K2.6: The 300-Agent Swarm
Moonshot AI released Kimi K2.6 on April 20, and it takes a fundamentally different approach to the “bigger is better” arms race. The model is a 1 trillion parameter MoE with 32 billion active parameters — competitive but not record-breaking on paper. What makes it unusual is the agent swarm architecture.
K2.6 can orchestrate 300 sub-agents across 4,000 coordinated steps in a single session. Its predecessor, K2.5, maxed out at 100 sub-agents and 1,500 steps. The model can run autonomously for up to 13 hours, decomposing complex engineering tasks into parallel workstreams, delegating to specialized sub-agents, and reassembling the results.
On SWE-Bench Pro, K2.6 posts a 58.6 — just above GLM-5.1 and tying with GPT-5.5. It does this at roughly a quarter of Claude Opus’s API cost. The model is released under a Modified MIT license, with weights on Hugging Face and available through Ollama for local deployment.
The agent swarm approach matters because it shifts the bottleneck. Instead of asking “how smart is the model?” the question becomes “how well can it coordinate work?” K2.6 suggests that orchestration capability may matter more than raw intelligence for real engineering tasks.
The Three-Month Gap Is Now Official
Epoch AI, the research organization that tracks AI capabilities with the rigor of an actuary, has updated its analysis: open-weight models now lag proprietary frontier models by approximately three months on their Epoch Capabilities Index.
Three months. Down from 12 to 18 months in 2023.
The gap varies — it occasionally closes to zero when a strong open release lands before the next proprietary jump, then widens again when a new GPT or Claude ships. But the trend line is unmistakable: the structural advantage that closed-source models once held is compressing. AgentBreaking’s analysis noted that Kimi K2.6, GLM-5.1, and DeepSeek V4 Pro have collectively “closed the gap on closed-source frontier models in ways that matter for actual work: multi-step task completion, tool call accuracy, and recoverable failure modes.”
That last point deserves emphasis. Benchmarks measure peak performance on controlled tasks. What matters in production is whether a model can recover when things go wrong — retry failed API calls, adjust its approach when code doesn’t compile, handle edge cases it wasn’t specifically trained on. The latest open-weight models are increasingly competitive on exactly these unglamorous-but-essential capabilities.
Robotics Goes Open Source
The open-source movement is spreading beyond language models into physical AI. Two developments this week illustrate the trend.
Hugging Face’s LeRobot v0.5.0 added support for its first humanoid robot (the Unitree G1), autoregressive vision-language-action policies, and faster datasets. The LeRobot paper was accepted to ICLR 2026. But the adoption numbers tell the bigger story: according to Hugging Face’s spring 2026 report, robotics datasets on the Hub grew from 1,145 in 2024 to 26,991 in 2025 — a 23x increase that made robotics the single largest dataset category on the platform, surpassing even text generation.
NVIDIA, meanwhile, open-sourced Isaac GR00T N1.6, a foundation model for humanoid robots that integrates vision-language-action policies with reasoning capabilities from its Cosmos platform. The model handles full-body loco-manipulation tasks — walking, reaching, grasping, coordinating — and is available on Hugging Face. The newer N1.7 builds on this with 20,000 hours of human video pretraining data, transferring manipulation skills learned from watching humans directly to robot control.
Open-source robotics models are following the same trajectory that language models did: start niche, get good fast, and become the default infrastructure that everyone builds on.
What This Means
The argument about whether open-weight models can match proprietary ones is effectively over for most production use cases. GLM-5.1 sits at #1 on the hardest coding benchmark. Kimi K2.6 ties GPT-5.5 at a fraction of the cost. The Epoch data confirms the gap is three months and shrinking.
The remaining advantage for proprietary models is diminishing to specific niches: the absolute cutting edge of novel reasoning, the first few months after a new capability breakthrough, and the convenience of managed APIs. For everything else — and that’s most real-world work — an MIT-licensed model you can run on your own hardware is either at parity or close enough that the cost and control advantages of open weights tip the decision.
The next frontier is physical AI, and it’s already going open source before the proprietary players can lock it down. When NVIDIA and Hugging Face are racing to open-source robot foundation models, the direction is clear.