Ant Group Open-Sources Ling-3.0-tiny: 1.3B Active Params Beats 31B Models
智东西 · wechat · 2026-08-11
Ant Group's Bailian team has open-sourced Ling-3.0-tiny, a lightweight hybrid reasoning MoE model designed for efficient local deployment. It has 7.9B total parameters but only activates 1.3B during inference.
- Architecture: Uses a 3:1 alternating stacking of KDA (Kimi Delta Attention) and MLA (DeepSeek's Multi-head Latent Attention), combined with 128 sparse experts to balance long-context capabilities and compute costs.
- Performance: Scores 25 on the ArtificialAnalysis index, closely trailing Gemma-4-26B, and surpasses Gemma-4-31B in agentic tasks.
- Edge Deployment: Offers BF16, FP8, and INT4 versions. Verified on DGX Spark, MacBook, and Mac mini. The FP8 version achieves 86-90 tokens/s on M4 Pro MacBooks with first-token latency under 100ms.
More from coding & agent
- Connected steps are not a workflow: nobody designs the handoff — usehive · 2026-10-16
- Making Blender animations with Opus, Astra and Argon: how you prompt still matters — chongdashu · 2026-10-16
- Running PolyStrata + Cline coding agent on a $0-fancy laptop at ~90 tok/s — rpwoerk · 2026-10-16
- How do you attribute agent gains: model training or infrastructure? — jrxna · 2026-10-16
- Feed Claude your books, blogs and top engineers' code for claude.md — 'it now writes better code than me' — blaizedsouza · 2026-10-16
- Open-source /label plugin auto-names and colors Claude Code sessions in one step — hermanschutte · 2026-10-16