Paper Instruments launched Ultramarine 1 Flash for professional knowledge work (finance, law, consulting, ops)—API at $0.15/$0.50 per 1M in/out, claims Fable-class JobBench 62.9 and ~60x cheaper than Fable per task; delivers Excel/PPT/Word and ERP/MES workflows; weights coming soon on Hugging Face, training code MIT.
Tencent Hunyuan open-weights Hy-Memory-30B-A3B (~3B active MoE) to replace the general LLM on the write path of Mem0/MemOS/Hy-Memory—trained with on-policy distillation + attributed memory RL; beats Claude Opus 5 / DeepSeek V4.1 Flash on a fixed cross-harness memory-writer matrix; Apache-friendly HF weights + GitHub framework.