The latest wave of open-weight model releases shows a field far more crowded and dynamic than many analysts predicted. New entrants and established labs alike are shipping capable models under varied licenses, challenging assumptions about consolidation in the industry.
Thinking Machines and Tencent Make Strategic Moves
Thinking Machines, a company announced in February 2025, has released Inkling, its first open model. The 975B-A41B multimodal mixture-of-experts model supports text, images, and audio as inputs, with text as output. It is not the strongest model among peers in China in its size class, but it is positioned as a base for fine-tuning via the company's Tinker service.
The company also released a smaller version, the 276B-A12B. This dual release signals a strategy of offering scale options to developers. Thinking Machines was not expected to be an open models company, yet it is now releasing the best open-weight models built in the USA, ahead of early leaders NVIDIA and Arcee.
The company's open model fine-tuning service generates hundreds of millions in revenue per year. That figure underscores the commercial viability of open approaches. The total effort for strong models now requires hundreds of millions to billions of dollars in training investment, a cost that once seemed to guarantee consolidation.
Tencent released Hy3, a 295B-A21B mixture-of-experts model that improves over its predecessor across all metrics. The previous Tencent version used a custom restrictive license. Hy3 switches to Apache 2, a notable shift toward openness. The model also proved a 50-year-old math problem using a dedicated harness and Sol as a judge, though it remains unclear how important Sol was to that result.
DeepSeek, Moonshot, and Poolside Push Boundaries
DeepSeek released DeepSeek-V4-Flash-0731 one day after OpenAI dropped prices of its smallest model by 80%. The Flash model beats Luna at the Pareto frontier, while the bigger DeepSeek model has not been updated yet. For the initial V4 releases, Flash was the star in performance per parameter, while Pro underwhelmed.
Moonshot AI released Kimi K3, the biggest open model release in some time, under a noncommercial license. The license requires inference and fine-tuning providers to enter a commercial agreement. Commentators Kevin Xu and Graham Webster co-authored a post about the license implications, questioning how much revenue-share licenses like Kimi K3 can stick. The license could also enable future government action against US entities doing business with Chinese AI companies.
Poolside made its third appearance in three consecutive months with Laguna-S-2.1 and Laguna-XS-2.1. The larger model, a newly pre- and post-trained version of the 118B-A8B MoE, fits on DGX Spark hardware. The smaller update is a 33B-A3B MoE. Poolside adopted the OpenMDW license, which is Apache 2.0-like with better legal backing for AI models. The company's blog includes evaluation trajectories, a transparency practice that stands out in the field.
Hardware Milestones and Global Lab Entries
Stay ahead of the AI curve
The most important updates, news, and content — delivered weekly.
No spam. Unsubscribe anytime.
Meituan LongCat, described as the Chinese DoorDash, released LongCat-2.0, a 1.6T parameter mixture-of-experts model trained entirely on Ascend 910s. This is the first non-Huawei, non-toy model trained entirely on Chinese accelerators. Other Chinese chips have mostly been used for inference, making this a significant milestone for training on domestic hardware.
AMD released Instella-MoE-16B-A3B-Think, a 16B-A3B MoE trained on Instinct cards. The company provides base and SFT checkpoints, MidTrain, and DPO stages, giving developers a full stack of resources. This entry from a semiconductor giant adds another layer to the open model ecosystem.
Motif Technologies, a Korean AI company, released Motif-3-Beta, its most ambitious model yet. The 314B-A13B MoE introduces architectural innovations called GDLA and mHC, pushing the technical envelope in new directions.
Swiss AI released Apertus-v1.5-70B, a continued pre-train of the fully open-source Apertus 1.0 using 2T more tokens. This approach of building on existing open models rather than starting from scratch is becoming more common.
The Consolidation Question Revisited
In 2024, the author predicted consolidation would pick up in 2026 or 2027. Training costs increasing by orders of magnitude every year made consolidation seem inevitable. Yet more companies are training strong models than expected, and an increasing number of organizations are releasing models openly.
Chinese labs are sustaining their pace of releases, and Xiaomi is a newer entrant accumulating mindshare in the AI economy. The proliferation of capacity to train strong models suggests the consolidation prediction may be wrong. A safer bet is to predict continued adoption of open models.
Demand for tokens is incredibly high and likely to increase. Building token machines is a likely path to value. The open models' role in the AI economy is the key question, and revenue-share licenses like Kimi K3 are a test case for how that role might evolve.
We're entering the decisive era. The question of how much market share open models can take remains open, but the momentum behind them is undeniable. Whether through license changes like Tencent's move to Apache 2, or through hardware milestones like LongCat-2.0's training on Ascend 910s, the open model movement is expanding in scope and ambition.
The article was published on Aug 02, 2026, as part of the Latest open artifacts series, and reflects a field in rapid motion. With new players entering monthly and established labs shifting strategies, the only certainty is continued change.