Open-Source AI

Today's Stories

1. 1

   Prism ML shrinks Qwen3.8 into 5.9GB Bonsai 2 27B

   Prism ML Inc. launched Bonsai 2 27B, which compresses a Qwen3.8 27B-based model from about 56 gigabytes to about 5.9 gigabytes while keeping around 98.2% of its capabilities. The model claims to hold onto most of its original abilities at a fraction of the size. That could let users run capable AI locally instead of sending data to a cloud service.

   The gains hinge on whether benchmark parity holds in real-world tasks, not just test scores. Watch the agentic and tool calling gap of three points against the original model.

2. 2

   HarnessRouter standardizes agent runs via one protocol

   HarnessRouter, an open-source implementation of the Unified Harness Protocol, runs Codex and Claude Code through one Responses-style API, currently at version 2026-09-12. Teams can pick or swap the agent runtime per feature without rebuilding progress, session, file, and cancellation handling in their product.

   UHP standardizes the task lifecycle but does not make harnesses behave the same, so identical prompts can still produce different plans and tools. Local sessions use separate OS users, not separate containers.

3. 3

   Kimi K3 comes to Amazon Bedrock

   Moonshot AI's Kimi K3 is now available on Amazon Bedrock. Kimi K3 is the first open-weight model on Bedrock with explicit prompt caching. The aim is to cut latency and input costs when reusing context across calls.

   The value hinges on whether prompt-reuse workloads actually hit the cache. Cached tokens are charged at a higher upfront rate and are retained for at least 30 minutes.

4. 4

   Kadai concept to be presented by two contributors at KDE conference

   KDE's annual Akademy conference opens September 19 at Graz University of Technology, where Eva Brucherseifer and Jan Muehlig will present the "Kadai" concept. Building on the pair's 2003 usability study of KDE 3.1, they argue that personal AI has reached the point where it can assemble the desktop on a per-user basis.

   The proposal is seen as potentially polarizing the room, while the trajectory of the Trinity successor and progress on the Sonic Desktop Environment will also be focal points.

5. 5

   Nvidia duo take open-vs-closed AI debate to Disrupt 2026

   Nvidia's Nader Khalil and Sydney Sykes will lead 'The Open vs. Closed AI Debate Is Just Getting Started' on the Builders Stage at TechCrunch Disrupt 2026, October 13-15 in San Francisco. The session treats the choice between open and proprietary models as a business decision affecting cost, infrastructure, margins, differentiation, speed, and control, not a philosophical argument.

   The debate will likely hinge on whether either approach can deliver a lasting competitive advantage as Nvidia CEO Jensen Huang argues the future is 'proprietary and open' rather than one or the other.

6. 6

   Hugging Face Skills tame Kiro and Claude Code on SageMaker AI

   Six open-source skills from the Hugging Face Skills GitHub repo let coding agents like Kiro and Claude Code deploy a Hugging Face model to Amazon SageMaker AI, automatically picking vLLM, autoscaling, and CloudWatch alarms. Unguided agents initially failed on this same task, so the skills are meant to prevent the fragile, costly endpoints that result when an agent lacks current deployment knowledge.

   The endpoint still bills continuously at $1.408/hr per instance whether it serves traffic or not, so teams must run the provided teardown script and confirm the resources are gone.