Version 5

Three open-weights models released in the week of June 18–26, 2026 have narrowed capability gaps with closed frontier systems on coding, math, and abstract reasoning benchmarks. GLM-5.2 (Zhipu AI) leads open-weights models on SWE-Bench, holds a verified 22.8% on ARC-AGI-2 at $0.25 per task[11], and topped PostTrainBench at 34.29%[13]; it has since attracted mainstream tech press coverage[16] and active developer community discussion[17]. VibeThinker-3B (Weibo AI) reports math and coding scores approaching Claude Opus 4.5 from only 3 billion parameters[3]. Ornith-1.0 (DeepReinforce) scores 82.4 on SWE-Bench Verified under MIT license[4], and Simon Willison's local testing of its 35B MoE variant confirms proficient agentic performance at 103 tokens per second[5].

Open-weights models reliable enough for independent capability auditing — including as calibration references for estimating closed model performance — offer developers a cheaper, inspectable alternative to proprietary APIs. The distillation question remains the key caveat: if these models' scores partly reflect training on closed-source outputs rather than independent capability, both the benchmark rankings and any calibration estimates built on them may overstate generalization.

Three open-weights models released in mid-to-late June 2026 have produced benchmark results that narrow the gap to closed frontier systems on software engineering, mathematical reasoning, and abstract visual tasks. GLM-5.2 from Zhipu AI leads open-weights models on the DeepSWE software engineering leaderboard[1] and scores near Claude Opus 4.7 on traditional LLM benchmarks[2]. VibeThinker-3B, released by Weibo AI and built on Qwen2.5-Coder-3B using a post-training pipeline called Spectrum-to-Signal combining supervised fine-tuning and GRPO, reports 94.3 on AIME26 and 80.2 Pass@1 on LiveCodeBench v6 — numbers that approach Claude Opus 4.5 from only 3 billion parameters, under MIT license[3]. Ornith-1.0-397B from DeepReinforce is a Mixture-of-Experts model with 397 billion total parameters and only 17 billion active, scoring 82.4 on SWE-Bench Verified under MIT license[4]; it comes in four variants — 9B Dense, 31B Dense, 35B MoE, and 397B MoE — all built on Gemma 4 and Qwen 3.5 under Apache 2.0 licenses[5].

The primary analytical debate centers on GLM-5.2. Zvi Mowshowitz has offered two framings in tension: his initial analysis called it 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[2]. A 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[6]. Community evidence leans toward the optimistic view: a normally benchmark-skeptical observer found GLM-5.2 genuinely impressive in direct use[7], a developer claimed it outperformed Opus 4.8 in a live coding build-off[8], multiple developers report real cost savings using it in Cursor[9], and at least one developer confirms that DeepSWE benchmark results match their own hands-on experience[10]. On abstract visual reasoning, GLM-5.2 scored a verified 22.8% on ARC-AGI-2 at $0.25 per task[11] — the highest-scoring open-weights model publicly reported on this benchmark, though well below GPT-5.5's leading 85%[12]. It also topped PostTrainBench — which tests autonomous training data selection, code writing, and model fine-tuning within a 10-hour single-GPU constraint — at 34.29%, against a 51.14% human-built instruct model baseline[13].

Simon Willison tested Ornith-1.0's 35B MoE GGUF locally and reported that it runs the agent harness across many tool calls proficiently at 103 tokens per second, with "very good" initial impressions. He also noted that the Gemma 4 base carries Apache 2.0 licensing without the restrictive additional terms that had complicated earlier Gemma model deployments[5]. A widely-shared analysis has used GLM-5.2's officially verified ARC-AGI-2 score as a calibration anchor, combined with WeirdML benchmark results, to assemble 8 private benchmark datapoints and triangulate estimates of closed model capabilities[14][15]. This approach is logically vulnerable to the same distillation concern Mowshowitz raises: if GLM-5.2's scores partly reflect training on closed-model outputs rather than independent capability, calibration estimates built on them may be systematically off. The cumulative picture is that open-weights models have made measurable advances on agentic coding, math, and abstract reasoning, but gaps to leading closed models remain substantial on the most demanding benchmarks, and the distillation question is unresolved for all three models.