Version 4

Three open-weights models released in the week of June 18–26, 2026 have narrowed capability gaps with closed frontier systems on key benchmarks. GLM-5.2 (Zhipu AI) leads open-weights models on SWE-Bench, scored 22.8% on ARC-AGI-2 — officially verified by the ARC-AGI evaluation team[10] — and topped PostTrainBench at 34.29%[16]. VibeThinker-3B (Weibo AI) reports math and coding scores approaching Claude Opus 4.5 from only 3 billion parameters[3]. Ornith-1.0-397B (DeepReinforce) scores 82.4 on SWE-Bench Verified under MIT license[4]. A widely-shared analysis uses GLM-5.2's verified ARC-AGI score as a calibration anchor to estimate closed model capabilities from 8 private benchmark datapoints[12][13].

Open-weights models reliable enough to serve as calibration references for closed-model capability estimation offer independent capability auditing, not just cheaper API alternatives. The distillation question remains the key caveat: models trained on closed-source outputs may produce misleading calibration signals precisely on tasks where benchmark numbers overstate generalization.

Three open-weights models released in the week of June 18–26, 2026 have advanced the argument that freely available AI can approach closed frontier performance on key benchmarks. GLM-5.2 from Zhipu AI scores near Claude Opus 4.7 on traditional LLM benchmarks[1] and leads open-weights models on the DeepSWE software engineering leaderboard[2]. VibeThinker-3B, released by Weibo AI (Sina Weibo's AI subsidiary), is a 3-billion-parameter dense model built on Qwen2.5-Coder-3B using a post-training pipeline called Spectrum-to-Signal combining supervised fine-tuning and GRPO; its reported scores — 94.3 on AIME26 and 80.2 Pass@1 on LiveCodeBench v6 — approach Claude Opus 4.5, a model estimated to be orders of magnitude larger, and the model is MIT-licensed[3]. Ornith-1.0-397B, released June 25 by DeepReinforce, is a Mixture-of-Experts model with 397 billion total parameters but only 17 billion active, scoring 82.4 on SWE-Bench Verified under MIT license[4].

The primary analytical debate surrounding GLM-5.2 comes from Zvi Mowshowitz, who has offered two framings in tension. His initial analysis called GLM-5.2 the best available open-weights model but argued it was heavily distilled from Claude Opus — noting its tendency to self-identify as Claude and its reliance on a Claude-based harness — and concluded it occupied a commercially awkward position: not cheap enough for bulk tasks, not capable enough for the hardest ones[1]. His subsequent newsletter elevated this to a potential 'DeepSeek moment' for open-source agents, suggesting frontier agentic capabilities may now be accessible in open weights for the first time[5]. Community testing has generally supported the optimistic view: a normally benchmark-skeptical observer found it genuinely impressive in direct use[6], a developer claimed it outperformed Opus 4.8 in a live coding build-off[7], multiple developers report real cost savings using GLM-5.2 in Cursor[8], and at least one developer confirms that DeepSWE benchmark results match their own hands-on experience[9].

On abstract visual reasoning, GLM-5.2 scored 22.8% on ARC-AGI-2 at $0.25 per task, a result that received official verification from the ARC-AGI evaluation team[10][11]. This places it well below GPT-5.5's leading 85% but as the highest-scoring open-weights model publicly reported on ARC-AGI-2. The benchmark's rapid progression — from a 3.0% best score in May 2025 to 85% by June 2026 — has prompted analysts to ask whether it still discriminates at the top, though the 62-point gap between GLM-5.2 and GPT-5.5 suggests meaningful separation persists between open and closed systems on this test. A widely-shared analysis combining GLM-5.2's officially verified ARC-AGI score with WeirdML benchmark results assembled 8 private benchmark datapoints, using GLM-5.2 as a calibration reference to triangulate estimates of closed model capabilities[12][13][14][15].

On June 26, GLM-5.2 also topped PostTrainBench — which tests whether an AI agent can autonomously select training data, write training code, run fine-tuning experiments, and submit an improved model within a 10-hour single-GPU constraint — at 34.29%, against a 51.14% human-built instruct model baseline[16]. Zhipu AI's official Hugging Face blog frames GLM-5.2 as built specifically for long-horizon tasks[17]. The cumulative picture is that open-weights models have made measurable advances on agentic coding, mathematical reasoning, and abstract visual tasks, but gaps to leading closed models remain large on the most demanding benchmarks, and the distillation question means rankings may overstate out-of-distribution generalization for all three models.