DeepSeek released a design article and a prototype repo for a long-lived router agent that hands work to isolated workers triggered by GitHub events.
Why it matters: Companies building multi-agent coding workflows get a concrete, open reference for routing work across parallel workers, but the stated gaps in memory and security mean it is not yet production ready.
1 source
ISTA DAS Lab's experimental build keeps half the experts per layer and uses low-bit quantization, reporting 98.7% and 94.8% of base benchmark scores.
Why it matters: Pruned and quantized large coding models could fit on smaller hardware, but the reproducibility questions mean buyers should verify results before relying on them.
1 source
At APSARA, Qwen's head Liu Dayi Heng reportedly said Qwen models will remain open source, countering rumors of API-only releases.
Why it matters: Open-weight availability from Qwen affects self-hosting and vendor-lock-in decisions for companies planning long-term model stacks.
1 source
The open-source Heretic 2.0 replaces rank-1 abliteration with an optimization method, adds a plugin system and breaks compatibility with 1.x configs.
Why it matters: Teams that experiment with editing open-weight models get a more flexible toolkit, but must migrate configs and should review the model behavior changes they introduce.
1 source