Version 6

Four open-weights models released between mid-June and early July 2026 have produced benchmark results narrowing capability gaps with closed frontier systems. GLM-5.2 (Zhipu AI) leads open-weights models on SWE-Bench and holds a verified 22.8% on ARC-AGI-2[17]; VibeThinker-3B reports math and coding scores approaching Claude Opus 4.5 from 3 billion parameters[3]; Ornith-1.0 (DeepReinforce) scores 82.4 on SWE-Bench Verified under MIT license[4]. A new entrant, Agents-A1 (35B), claims 1-trillion-parameter-model performance by training on long verified multi-step task trajectories and distilling specialist teacher models rather than scaling parameters[14].

Open-weights models reliable enough for independent capability auditing offer developers cheaper, inspectable alternatives to proprietary APIs. The distillation question remains the central caveat: if benchmark scores partly reflect training on closed-source outputs rather than independent capability, both rankings and any calibration estimates built on them may overstate generalization.

Three open-weights models released in mid-to-late June 2026 — GLM-5.2 from Zhipu AI, VibeThinker-3B from Weibo AI, and Ornith-1.0 from DeepReinforce — have produced benchmark results that narrow the gap to closed frontier systems on software engineering, mathematical reasoning, and abstract visual tasks. GLM-5.2 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, built on Qwen2.5-Coder-3B using a post-training pipeline 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 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]; Simon Willison tested its 35B MoE GGUF locally and reported proficient agentic performance at 103 tokens per second, with clean Apache 2.0 licensing from its Gemma 4 base[5].

The primary analytical debate centers on GLM-5.2. Zvi Mowshowitz initially 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]. He later elevated this to a potential 'DeepSeek moment' for open-source agents, without resolving the distillation concern[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 — the highest-scoring open-weights model publicly reported on this benchmark, well below GPT-5.5's 85%[11]. It also topped PostTrainBench at 34.29%, against a 51.14% human-built instruct model baseline[12]. A circulating discussion frames GLM-5.2's token volume on DeepSWE as potentially cost-inefficient despite its benchmark wins, adding a dimension to the commercial viability debate Mowshowitz had flagged[13].

A fourth model, Agents-A1 (35B), takes a different approach to the efficiency question. Rather than scaling parameters or optimizing post-training pipelines, it trains on long verified multi-step task trajectories averaging 45K tokens — teaching the model complete work processes including search, tool use, error correction, and answer verification. Specialist teacher models are trained separately for search, science, instruction following, and tool use, then their capabilities are distilled into a single student model. The result claims strong performance on long-task benchmarks across all four domains, under Apache 2.0 license[14]. This is a different architectural bet than GLM-5.2's or Ornith-1.0's approaches: rather than scale MoE or tune post-training, it tries to compress the behavioral patterns of specialized frontier systems into a compact model through richer trajectory data.

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[15][16]. 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, the distillation question is unresolved across multiple models, and cost-per-task efficiency of high-performing open models is emerging as a secondary concern.