Open-Source & Specialized Models (MiMo, BU-30B, MiMo-V2-Flash)

Multiple vendors released or detailed open and specialized models for long context and browser automation.
Key insights
Xiaomi publishes MiMo‑V2‑Flash focused on production efficiency
Xiaomi released MiMo‑V2‑Flash with a 309B parameter model that activates ~15B per token, native training at 32K and extended to 256K context, claimed 150 tokens/sec inference, and very low quoted costs ($0.10 per 1M input tokens; $0.30 per 1M output tokens). Xiaomi also published a demo chat web product to showcase the model's code abilities.
BU‑30B-A3B-Preview targets browser automation with MoE efficiency
browser_use released BU‑30B-A3B-Preview, a 30B‑parameter MoE-based model specialized for browser automation and DOM understanding. It supports visual-language input, 32k token context, and activates about 3B parameters at runtime to balance capability and resource cost. The model is pitched for web agents, RPA, and tasks requiring precise webpage element reasoning.
Hybrid SWA, MTP and MOPD engineering choices emphasize deployment pragmatism
Technical notes from Xiaomi and contributors highlight Hybrid SWA attention, fixed KV cache for deployment friendliness, multi-token prediction (MTP) layers for RL efficiency, and MOPD on-policy distillation to replicate RL teacher behavior at low compute. These choices prioritize production deployment, long context handling, and low-cost RL distillation.