Open Model Race Intensifies: Thinking Machines, Poolside, Moonshot Ship Competing Frontier Releases
A dense wave of open-weight model releases—including Thinking Machines' first model Inkling, Poolside's Laguna S2.1, and Moonshot AI's Kimi K3—signals that consolidation predictions for AI labs have not materialized. Licensing terms, particularly Kimi K3's noncommercial agreement requirement, are emerging as a new front in US-China AI policy.
The consolidation thesis is failing
Industry observers, including analysts at Interconnects, predicted in 2024 that rising training costs would force consolidation among AI labs by 2026-2027. That hasn't happened. Instead, more organizations are training frontier-capable models and releasing them openly, according to a roundup published by Interconnects covering the latest wave of open artifacts.
The clearest evidence: Thinking Machines, founded in February 2025, has released its first model. Few predicted the company would become an open-weights player—its open model fine-tuning service, Tinker, reportedly generates hundreds of millions of dollars in annual revenue, according to the report.
What shipped
Inkling (Thinking Machines): A 975B-parameter MoE model with 41B active parameters (975B-A41B), supporting text, image, and audio inputs with text-only output. Thinking Machines also released a smaller 276B-A12B variant. According to the report, Inkling is not the strongest model in its size class compared to Chinese competitors, but is positioned as a strong fine-tuning base via the company's Tinker service.
Laguna-S-2.1 (Poolside): A 118B-A8B MoE, small enough to run on a single DGX Spark. This marks Poolside's third consecutive monthly appearance in the Interconnects roundup. The company adopted the OpenMDW license—an Apache 2.0-style permissive license with stronger legal backing for AI models—and published full evaluation trajectories alongside the release.
Hy3 (Tencent): A 295B-A21B MoE that improves on its predecessor across all measured metrics. Tencent switched from a custom restrictive license to Apache 2.0 for this release, a notable shift from the prior version covered in Artifacts #21.
DeepSeek-V4-Flash-0731 (DeepSeek): An update to DeepSeek's V4 Flash model, released one day after OpenAI cut pricing on its smallest model by 80%. According to the report, the update beats a competing model called Luna on the Pareto frontier. The larger V4 Pro model has not yet received a corresponding update.
Kimi K3 (Moonshot AI): Described as the largest open model release in recent months. Kimi K3 ships under a noncommercial license requiring inference and fine-tuning providers to sign separate commercial agreements with Moonshot. Analysts Kevin Xu and Graham Webster argue this licensing structure could expose US companies using Kimi K3 commercially to new regulatory scrutiny, since a direct contractual relationship with a Chinese AI company creates a clearer target for policy restrictions than a standard open license would.
Other notable releases
- LongCat-2.0 (Meituan): A 1.6T-parameter MoE, notable as the first non-Huawei model of significant scale trained entirely on Chinese Ascend 910 accelerators rather than Nvidia hardware.
- Laguna-XS-2.1 (Poolside): A smaller 33B-A3B companion model.
- Motif-3-Beta (Motif Technologies): A 314B-A13B MoE from the Korean startup, introducing new architectural components referred to as GDLA and mHC.
- Apertus-v1.5-70B (Swiss AI): A continued pretraining run on top of the fully open-source Apertus 1.0, adding 2 trillion additional training tokens.
- Instella-MoE-16B-A3B-Think (AMD): A 16B-A3B MoE trained on AMD Instinct accelerators, released with full training checkpoints from base through SFT and DPO stages.
What this means
The volume and diversity of this release cycle—spanning a well-funded US startup, two Chinese tech giants, a Korean startup, a Swiss research consortium, and a chip vendor—undercuts the thesis that only a handful of labs can afford frontier-scale training. Capacity to train strong models is spreading, not concentrating.
The more consequential development is licensing, not capability. Tencent's move to Apache 2.0 for Hy3 lowers adoption friction, while Moonshot's noncommercial terms for Kimi K3 do the opposite—and could turn commercial usage into a geopolitical liability for US companies. As open models approach parity with closed frontier systems, license terms are becoming the primary lever labs use to control downstream value capture and, increasingly, a mechanism through which national policy could reach into private AI infrastructure decisions.
Related Articles
Open Model Race Intensifies: Thinking Machines, Tencent, Poolside, Moonshot Ship Frontier-Class Releases in Same Week
A wave of open-weight model releases from Thinking Machines, Tencent, Poolside, Moonshot AI, and Meituan signals that model-building capacity is spreading rather than consolidating. The releases range from a 1.6 trillion-parameter MoE trained entirely on Chinese accelerators to a noncommercial-licensed model raising new questions about US-China AI trade.
Chinese open-weight models claim top 5 spots on OpenRouter by usage, threatening US frontier labs
Chinese AI models from Tencent, Xiaomi, DeepSeek, MiniMax and Z.ai now occupy the top five positions by weekly token usage on OpenRouter, a major AI marketplace. The shift threatens to commoditize premium models from OpenAI and Anthropic as businesses reserve expensive frontier systems for only their hardest problems.
Moonshot AI's Free Kimi K3 Model Is Forcing OpenAI, Google, and Anthropic to Rethink Their Open-Weight Strategy
Chinese startup Moonshot AI released Kimi K3 as a free, open-weight model that it claims beats top US systems at a fraction of the cost. The move has intensified pressure on OpenAI, Google, and Anthropic to reconsider their closed-model strategies.
Chinese AI Models Capture 30%+ of U.S. Developer Token Usage as OpenAI, Anthropic Costs Rise
Chinese AI models including DeepSeek and Z.ai have captured over 30% of weekly token usage by U.S. companies on OpenRouter since February 2025, up from 4.5% in the first half of the year. The shift comes as companies seek alternatives 60-90% cheaper than leading models from OpenAI and Anthropic, while Chinese models close the performance gap to within 6-9 months of U.S. frontier systems.
Comments
Loading...