Three Open-Weight Coding Models Rushed In as Fable 5 Went Dark

One-liner: The US Fable 5 export ban created an opening for open-weight models — and three of them shipped within days, collectively narrowing the gap that Anthropic's ban left for international developers.

Key Facts

• North Mini Code (Cohere, June 9 — shipped the same day as Fable 5 launch): 30B-parameter MoE with just 3B active params, Apache 2.0 license; delivers 2.8× higher output throughput than Devstral Small 2 on SWE-Bench
• Kimi K2.7-Code (Moonshot AI, June 12 — same day as the ban): 1-trillion-parameter MoE, Modified MIT, 256K context, priced at $0.95/M tokens; community benchmarks place it above Claude Opus 4.8 on agentic coding tasks
• GLM-5.2 (Zhipu AI, June 13): MIT license, 1M-token context window; AI analysis community called it "the real deal" on real-world coding tasks
• Z.ai forecasters now give 70%+ odds that an open-weight model matching Fable 5 capability arrives before year-end 2026

Why It Matters

Fable 5 held the highest community score on agentic coding benchmarks when it launched — and was pulled from international users three days later. The rapid response from Chinese AI labs and Cohere shows that export controls on frontier models accelerate open-weight development rather than suppress it. For enterprises outside the US, the calculus has shifted: self-hostable models are now within striking distance of proprietary frontier performance on software tasks, and they carry no geopolitical supply-chain risk.

More

• Kimi K2.7-Code release notes — explainx.ai
• Fable 5 ban: what open weights filled the gap — The New Stack
• GLM-5.2 community analysis — Latent Space