Latest open artifacts (#23): Laguna S2.1, Inkling, & Kimi K3 show the utility of open models on the Pareto frontier
Industry consolidation predicted for 2026-2027 has not materialized; instead, more companies are investing hundreds of millions to billions in training strong open models Token demand is surging as models become more efficient, making "building token machines" a recognized path to value for labs Thinking Machines emerged as an unexpected open-model leader, with their finetuning service generating hundreds of millions in annual revenue while releasing top U.S. open-weight models Chinese labs main
Analysis
TL;DR
- Industry consolidation predicted for 2026-2027 has not materialized; instead, more companies are investing hundreds of millions to billions in training strong open models
- Token demand is surging as models become more efficient, making "building token machines" a recognized path to value for labs
- Thinking Machines emerged as an unexpected open-model leader, with their finetuning service generating hundreds of millions in annual revenue while releasing top U.S. open-weight models
- Chinese labs maintain sustained momentum with new entrants like Xiaomi and releases from Tencent, Moonshot AI, and Meituan, complicating the open vs. closed model landscape
- Licensing strategies are diverging: Apache 2.0 (Tencent, Poolside), OpenMDW (Poolside), and revenue-share/noncommercial licenses (Kimi K3) reflect competing visions for open model commercialization
Why It Matters
The failure of consolidation predictions to materialize signals a more competitive and fragmented AI ecosystem than anticipated, with open models carving out significant commercial and technical space. For practitioners, this means open-weight models are no longer second-class citizens—they are competitive, well-licensed, and commercially viable, reshaping deployment and fine-tuning strategies. The licensing diversity (permissive vs. restrictive) also introduces strategic considerations for enterprises choosing which models to adopt.
Technical Details
- Inkling by Thinking Machines: A 975B-A41B multimodal MoE supporting text, image, and audio inputs with text output; a smaller 276B-A12B variant is also released and noted as highly competitive for its size. Positioned as a fine-tuning base via the Tinker commercial service.
- Hy3 by Tencent: A 295B-A21B MoE improving over its predecessor across all metrics; notably switched from a restrictive custom license to Apache 2.0. Demonstrated ability to prove a 50-year-old math problem.
- Laguna-S-2.1 by Poolside: An 118B-A8B MoE that fits on a single DGX Spark, newly pre- and post-trained. Released under the OpenMDW license (Apache 2.0-like with stronger AI-specific legal backing). Full evaluation trajectories published transparently.
- DeepSeek-V4-Flash-0731 by DeepSeek-AI: Released one day after OpenAI cut its smallest model's prices by 80%; beats Luna at the pareto frontier in performance per parameter. The larger Pro variant was underwhelming in initial V4 releases.
- Kimi-K3 by Moonshot AI: Described as the biggest open model release in some time, released under a noncommercial license requiring commercial agreements for inference and fine-tuning providers. Raises policy questions about U.S.-China AI business relationships.
- LongCat-2.0 by Meituan: A 1.6T-parameter MoE from the Chinese "DoorDash"; strong on benchmarks but not the most capable for its size class beyond benchmark performance.
Industry Insight
- The open-model ecosystem is maturing faster than expected, with companies like Thinking Machines proving that open-weight releases can generate hundreds of millions in revenue through finetuning services—challenging the assumption that only closed models are commercially viable.
- Licensing is becoming a strategic differentiator: permissive licenses (Apache 2.0, OpenMDW) attract developer adoption, while restrictive licenses (Kimi K3) attempt to capture commercial value and navigate geopolitical risk, creating a fragmented landscape enterprises must navigate carefully.
- Chinese labs continue to compete aggressively on both open and closed fronts, suggesting that U.S.-centric consolidation narratives may underestimate the global pace of open-model development and the role of revenue-share licensing as a middle ground.
Disclaimer: The above content is generated by AI and is for reference only.