Open Models Ecosystem Thrives: Interconnects Covers Kimi K3, DeepSeek & More
Interconnects (Nathan Lambert) · rss · 2026-08-02
In a comprehensive recap of open models, Nathan Lambert highlights that despite predictions of industry consolidation due to rising training costs, the open-source ecosystem is thriving. Companies are realizing that building token machines is a viable path to value amid surging demand.
Key releases covered include:
- Thinking Machines (Inkling): A 975B-A41B multimodal MoE model. While not the absolute strongest in its class, it serves as an excellent base for fine-tuning, driving significant revenue for their commercial tuning service.
- Tencent (Hy3): A 295B-A21B MoE model that switches to the permissive Apache 2.0 license and outperforms its predecessor across all metrics.
- Poolside (Laguna S2.1): An update to their MoE model adopting the AI-friendly OpenMDW license, offering exceptional transparency with full evaluation trajectories.
- DeepSeek (V4-Flash-0731): Released just a day after OpenAI slashed prices, beating competitors on the Pareto frontier of performance per parameter.
- Moonshot (Kimi K3): A major release under a non-commercial license requiring commercial agreements for inference and fine-tuning providers. The author notes this could potentially trigger future US government regulatory actions.
- Meituan (LongCat-2.0): A massive 1.6T parameter MoE notable for being trained entirely on Huawei Ascend 910 chips, marking a first for non-Huawei Chinese accelerators.
More from Models
- Recent LLM Math Successes Suggest Formal Methods Aren't Strictly Required — littmath · 2026-08-02
- Microsoft's First Native Real-Time Bidirectional Voice Model MAI Realtime Surfaces — testingcatalog · 2026-08-02
- 8B Model Beats Frontier LLMs: Replit CEO's Chess Engine Hits 1500 Elo — amasad · 2026-08-02
- Test: AI Models Hallucinate Missing Tool Parameters Like Emails — eliobldr · 2026-08-02
- Vertical AI Companies Accelerate: Open-Source Fine-Tuning and Routing Optimize Costs — lqiao · 2026-08-02
- Over a Third of New Web Text is AI-Generated; Pangram 4 Claims 1-in-24k False Positive Rate — aftahi_ai · 2026-08-02