Open-Source AI

Today's Stories

1. 1

   Together AI raises $800M at $8.3B valuation on strong open-source adoption

   Together AI, an AI infrastructure company founded in 2022, raised $800 million（約1300億円） in Series C funding at an $8.3 billion（約1.3兆円） valuation, led by Aramco Ventures with backing from Vista Equity Partners, General Catalyst, Emergence Capital, Nvidia, and others. The company last raised $305 million（約490億円） at $3.3 billion（約5300億円） valuation about 16 months ago. Together AI claims annual bookings of over $1.15 billion（約1800億円） as businesses increasingly adopt cheaper open-source AI models through neocloud providers—companies offering AI-specific hardware and infrastructure—rather than paying premium prices for closed proprietary models. The trend suggests a structural shift in how enterprises are choosing to build AI systems, moving toward cost-effective alternatives.

   Together AI says open-source model usage has tripled across the industry in the past year, and the company counts thousands of paying customers including Cursor, Cognition, and Decagon. The broader neocloud sector has become a hot investment target, with competitors like Upscale AI and TensorWave raising comparable funding rounds in recent weeks.

2. 2

   Alook launches open-source platform to coordinate AI agents as company teams

   Alook, an open-source self-hosted platform, lets users create multiple AI agents (such as Claude Code or OpenCode sessions) and organize them into an org chart with defined roles and reporting lines. Agents coordinate via email inboxes without manual intervention, and the article walks through building a four-agent price-tracking system that runs on a schedule unattended. The platform addresses a real coordination problem in AI agent systems—without structure, multiple agents can create chaotic interactions and lose context. By giving agents formal roles, reporting hierarchies, and asynchronous email-based communication (rather than direct message passing), Alook makes it feasible for small operators to build multi-agent workflows that work like a real team, potentially lowering the barrier for businesses to deploy AI automation.

   The platform detects whichever coding agent runtime is already installed on your machine and deploys agents there; it runs as a local daemon and opens a dashboard at http://localhost:15210. The GitHub repository is publicly available for setup and exploration.

3. 3

   Together AI raises $800M at $8.3B valuation to scale open-source AI

   Together AI, an infrastructure company for running open-source AI models, closed an $800 million（約1300億円） Series C funding round led by Aramco Ventures. The company's annual bookings crossed $1.15 billion（約1800億円） last quarter, and it now serves thousands of paying customers including Cursor, Cognition, and Decagon. Businesses using closed AI models face high costs that can consume their entire profit margin. Together AI offers an alternative by helping companies run open models like DeepSeek and Nemotron at a fraction of the cost—customers report cost savings of 6x to 60x compared to closed-model pricing for equal or better performance. Nearly three-quarters of organizations expect to increase their use of open-source AI, suggesting demand for affordable infrastructure continues to grow.

   Open-source model usage across the industry has tripled in twelve months. The company positions itself as the platform making open-source models genuinely usable at enterprise scale, serving many of the biggest names in AI.

4. 4

   Genesis AI model PEARL shows drug discovery can finally work—hitting real-world accuracy thresholds

   Genesis Molecular AI's PEARL model demonstrated on the OpenBind benchmark that it can accurately predict how small molecules bind to proteins, including modeling protein flexibility and induced-fit effects without fine-tuning on target-specific data. The model outperformed public competitors across evaluation metrics on 802 never-before-seen molecular complexes. Small-molecule drug discovery has long struggled because there are 10^60 drug-like molecules to search, and the properties that make a strong binder often conflict with those needed for the drug to reach its target in the body. PEARL's ability to model both ligand placement and protein adjustment together suggests that agentic drug-discovery loops—where AI iterates like a chemist, forming hypotheses and testing candidates—may now be practically feasible, potentially enabling 24/7 automated discovery cycles when paired with lab partners.

   The field has conventionally benchmarked poses at "2 Angstrom RMSD" accuracy, but Genesis argues that 1 Angstrom RMSD is the real threshold needed to correctly model molecular interactions like hydrogen bonds (which span only 0.6Å). PEARL's recent results suggest the community may be ready to abandon the weaker standard and pursue genuinely harder validation targets.

5. 5

   MOTHRAG open-sourced: graph-free multi-hop search beats knowledge-graph systems

   Researchers open-sourced MOTHRAG, a multi-hop retrieval framework that uses a dense index and query-time orchestration instead of offline knowledge graphs. On three benchmarks (HotpotQA, 2WikiMultiHopQA, MuSiQue), MOTHRAG scored 78.1 / 76.3 / 50.5 accuracy/F1, outperforming GraphRAG (68.6 / 58.6 / 38.5), HippoRAG (75.5 / 71.0 / 48.6), and RAPTOR (69.5 / 52.1 / 28.9). Graph-based systems require expensive offline indexing passes every time data changes—a constant cost for frequently updated corpora (prices, internal filings, support tickets, news). MOTHRAG avoids this by using only commodity APIs and simple embed-and-append updates, which may reduce operational overhead for businesses managing live data.

   The framework uses no knowledge graph and no GPU, keeping infrastructure simple. Updates operate as embed-and-append operations, lowering the indexing burden compared to full graph rebuilds.

6. 6

   Hugging Face and Cerebras demo real-time voice AI with low latency

   Hugging Face and Cerebras demonstrated a speech-to-speech AI pipeline that combines open-source models—Nvidia's Parakeet for speech recognition, Google DeepMind's Gemma 4 31B language model running on Cerebras inference, and Alibaba's Qwen3TTS for text-to-speech—to enable natural, fast conversational responses. The modular, open architecture allows developers to inspect, modify, and extend each component. Latency has been a major bottleneck in voice AI systems; many production systems experience multi-second delays at the P95 (worst-case scenarios), making conversations feel unreliable and unnatural. By making language-model inference dramatically faster and more stable, Cerebras addresses this bottleneck. For robots, voice assistants, and embodied AI, this responsiveness is not a cosmetic improvement—it makes interactions feel alive and natural at scale, which may enable more practical deployment of conversational AI in real-world applications.

   The pipeline already powers Reachy Mini robots, with more than 9,000 robots in the wild. Developers can explore the demo on Hugging Face Space and the code in the huggingface/speech-to-speech repository.