At 0.9B parameters, the smallest member of the K2 Horizon fleet is designed to bring capable, open AI to devices as compact and personal as a watch. Our goal is to make intelligence available at every scale—from everyday edge devices to enterprise clusters—while giving developers a common model family they can build and scale with. Small in footprint, but part of a much larger horizon. Meet K2 Horizon 0.9B.

🚀 ZDTaichu5.0-9B is now on ModelScope! 🤖 An on-device multimodal model from TaichuAI. At 9B parameters it runs on a single GPU and brings spatial reasoning, embodied AI and agentic tool use to edge deployment. Qwen3.5-9B backbone + C-RADIOv4-H vision encoder, 128K context, any-resolution image and video input. 🧭 Spatial reasoning: leads the compared 10B-scale open VLMs (Qwen3.5-9B, STEP3-VL-10B, gemma4-8B-E4B) and scores above Gemini 3 Pro, Grok 4 and GPT-5.2 on ViewSpatial, MMSI-Bench and MindCube-tiny 🛠️ Agent: highest among the compared open models on TAU2-Bench, Claw-Eval and IFEval 📄 First-tier results on documents, charts, OCR, visual math and video, with a ready-to-use vLLM branch and Docker image 🧠 Entropy-Gated Adaptive Recurrent Reasoning: extra latent refinement steps go only to the hard tokens

fx v0.0.9 Use a frontier model to steer, then hand implementation to a fast workhorse. • Give each subagent its own model and reasoning effort • Subagents keep working while you steer • Send feedback to subagents mid-task • Switch models with Ctrl+P without losing your draft or images • New conversations get short automatic titles • Better GitHub-style Markdown rendering • New libfx examples for Node.js, browsers, Next.js & Nuxt No new commands or concepts. Just chat with fx.

Base64z: Use the characters 0-9, A-Z, and a-y, as well as the digraphs z0, z1, and z2. This gives you a purely alphanumeric encoding, so it's double-click selectable in all contexts, without the decoding complexity of base62.