Open Source LLM model Brief — 2026-07-13

Top Stories

Edgen · 2026-07-13
Summary: Meituan has open-sourced its 1.6 trillion-parameter LLM, LongCat-2.0, under the permissive MIT license. Notably, the model was pre-trained entirely on over 50,000 domestic Chinese AI accelerators, marking a significant milestone in reducing dependency on Nvidia hardware for large-scale training. The model features a 1 million-token context window and is built on proprietary delivery-network data, differentiating it from general-purpose models .
Why It Matters: This release demonstrates that large-scale AI training is achievable on domestic hardware, potentially reshaping the global AI infrastructure strategy. For enterprises, it also highlights the growing value of proprietary, domain-specific data as a key differentiator, with Citi maintaining a Buy rating on Meituan, citing this open-source strategy as a “moat-builder” .
URL: Meituan open-sources trillion-parameter LLM as Citi keeps Buy rating

至顶网 (Zhiding) · 2026-07-13
Summary: Cohere has launched its first coding model, North Mini Code, under the Apache 2.0 license. The 30-billion-parameter MoE model (with 3 billion active parameters) is designed for agentic coding tasks and can run on a single H100 GPU. Cohere claims it outperforms Qwen3 and Gemma 4 on the Artificial Analysis coding index, furthering its “sovereign AI” philosophy by enabling local deployment .
Why It Matters: Cohere’s pivot to the developer market with a performant, open-weight coding model addresses the growing demand for self-hosted development tools. This move positions Cohere against other players like Mistral and JetBrains, offering enterprises a choice for building proprietary, secure, and cost-effective AI-powered developer toolchains .
URL: Cohere推出首个编程模型，从企业市场转战开发者

KuCoin · 2026-07-13
Summary: A German research consortium has released Soofi S 30B-A3B, an open-source LLM with 31.6 billion total parameters and 3.2 billion activated per token. Trained on a German AI cloud with 15.3% German-language data, it scored 73.8% on HumanEval and 84.2% on the German MBPP, outperforming models like OLMo 3 32B and Apertus 70B .
Why It Matters: The launch represents a concrete step toward European AI sovereignty, delivering a high-performing model optimized for the German language. Its efficiency and open-source nature provide a strong alternative for enterprises and governments prioritizing data security and local language compliance .
URL: German AI Consortium Launches Soofi S 30B-A3B, Outperforms on English and German Benchmarks

Iresearch.cn · 2026-07-13
Summary: Tencent has released the official version of its Hy3 model, which uses a MoE architecture with 295 billion total and 21 billion active parameters. The model is open-sourced under the Apache 2.0 license and is priced at 1 yuan per million input tokens. Hy3 shows a 20-30% improvement in agent and coding capabilities over its preview version and is being integrated across Tencent products .
Why It Matters: Hy3’s strong performance at a competitive price point reinforces the trend of high-quality, open-source models being commercially viable. Its integration into widely-used platforms like WeChat and WeGame validates the model’s real-world reliability and provides a blueprint for other large tech companies to leverage open-source AI .
URL: 腾讯混元Hy3正式版发布并开源，采用Apache 2.0协议

Yahoo Tech · 2026-07-13
Summary: The free, open-source GLM-5.2 from Z.ai has garnered attention from Silicon Valley developers for its coding and agentic capabilities. While early tests show it performs well on tasks like writing and research, users have noted significant capacity issues, slow response times, and flaws in feature reliability, such as design generation .
Why It Matters: Despite its current shortcomings, GLM-5.2 represents the next wave of highly capable, free AI models challenging the status quo. For cost-conscious users and enterprises, its potential to deliver comparable information to premium models for everyday tasks is a significant development, though reliability remains a major barrier to enterprise adoption .
URL: China’s free AI model is giving DeepSeek déjà vu. It works, but takes patience.