Open-Source AI

Today's Stories

1. 1

   Salesforce and Nvidia launch Koa reasoning model for enterprise AI

   Salesforce announced Koa, its first reasoning model, at its Dreamforce conference. It is built on Nvidia's open-weight Nemotron and post-trained by both companies for sales, marketing, and customer-support work. Koa is an open-weight alternative to closed frontier models, so tasks that previously routed to Claude or ChatGPT through Agentforce's gateway can now run on a model Salesforce says uses fewer tokens for the same work.

   Salesforce says it did not use any actual customer data to train Koa, instead crafting synthetic data that mimicked customer patterns. The test is whether that approach delivers on its efficiency and security claims.

2. 2

   SHADOW-50M: 44M model ships in 19.8 MB, ~1,900 tok/s on CPU

   The developer released SHADOW-50M, a 44M parameter model trained from scratch on 45B tokens. The 19.8 MB model uses ternary weights and a 73,880-token vocabulary, and runs offline. The previous SHADOW-250M was 60 MB and ran around 400 tok/s on CPU but could not reliably reason over retrieved data or calculate. SHADOW-50M targets those two gaps while cutting size and raising CPU speed.

   Since this is a proof of concept, not a product, the test is whether the calculation circuit approach holds up beyond the demo. Watch the WebAssembly browser version, which runs at around 500 tok/s.

3. 3

   InOrbit.AI ships OpenRobOps, first ISO 21423 reference implementation

   InOrbit.AI released OpenRobOps, an Apache 2.0 open-source robot fleet operations foundation it calls the industry's first ISO 21423 reference implementation. OEMs previously spent years building custom back ends and user interfaces; InOrbit says OpenRobOps lets teams bypass years of back-end development.

   The test is whether OEMs adopt ORO instead of building their own fleet managers, since InOrbit's paid Space Intelligence remains separate.

4. 4

   Mozilla: open models trail frontier AI by just 4.4 months

   Mozilla's State of Open Source AI report, published September 15, says the gap between top closed US frontier models and the best open-weights models has shrunk to 4.4 months. Moonshot AI's Kimi K3 lands just three points behind Anthropic's Fable 5 on the Artificial Analysis Intelligence Index at 30 percent of the cost, and DoorDash already runs Kimi for routine work.

   Krikorian frames the decision as workload-specific, not organization-specific — paying for closed buys about a four-month head start at roughly five times the per-task cost, and only for tasks taking 8 to 12 hours.

5. 5

   China's Guo Jiakun rejects Amodei's AI slowdown call

   At a September 14 press briefing, China's Foreign Ministry deputy spokesperson Guo Jiakun responded to a Reuters question about Amodei, Sam Altman and Elon Musk urging a slowdown in frontier AI, saying fears and vicious competition harm global AI governance. Amodei's essay "We Must Pace the Frontier", published on his personal site on the 12th, argued that pace coordination within democracies can only go as far as the US lead over China allows. He conditioned it on export bans on AI chips and semiconductor equipment to China.

   Guo did not name Amodei or any company directly, so the test is whether Beijing's push for a BRICS open-source AI community, which Xi Jinping proposed at the BRICS summit in India, gains traction. China chairs BRICS next year.

6. 6

   Agent Router joins Linux Foundation's AAIF as unified AI API

   The Linux Foundation's Agentic AI Foundation said the open-source Envoy AI Gateway has joined it and been renamed Agent Router, announced at AGNTCon＋MCPCon Japan 2026 on September 10. Instead of separate APIs for OpenAI, Anthropic and Google, Agent Router offers one OpenAI-compatible endpoint, so apps can switch AI models without rewriting code, and it already counts 11 public adopters including Bloomberg and Tencent Cloud.

   The Foundation says it will keep developing Agent Router and push it toward becoming a de facto industry standard, though no deadline is given. The project is already at version 1.1 and production-ready.