DeAI Daily Brief — 22 September 2026

Today in DeAI: Xiaomi open-weights a frontier-scale model family under MIT, Z.ai answers a privacy failure with an Apache 2.0 code drop, and Bloomberg reports open weights rescuing startup unit economics.

Xiaomi ships MiMo-V2.6 open weights, MIT-licensed, with the RL code

Xiaomi released the MiMo-V2.6 series on September 22: MiMo-V2.6-Pro-RL (524B parameters) and MiMo-V2.6-Flash-RL (159B) on Hugging Face under MIT, plus the reinforcement learning code, task environments, and technical report. Xiaomi claims a 46.32 Artificial Analysis Intelligence Index score and RL costs of about $2.62M for Pro — vendor figures, not independently reproduced. Why it matters: the training-code transparency is the new part; our full coverage separates the verifiable release from the claims.

Z.ai open-sources ZCode after workspace-upload disclosure

Z.ai open-sourced its ZCode coding agent under Apache 2.0 after a paying subscriber documented — with packet captures — the agent silently uploading roughly 313MB of workspace data, including .git histories, to an Alibaba Cloud bucket. Z.ai says the bucket was destroyed and nothing was retained or used for training; those assurances are claims, not verified facts. Why it matters: agent trust is a data-egress question, and open code is the only audit trail you control. Our earlier piece on zero-data-retention API claims covers the same trust boundary from the API side.

Bloomberg: Harvey, Abridge, Ramp move workloads onto open weights

Bloomberg reports that several well-funded AI startups shifted workloads off frontier-lab APIs after usage-based pricing pushed margins negative, post-training in-house on open-weight bases including Moonshot AI's Kimi K3. The margin figures are Bloomberg-reported company claims, not audited numbers. Why it matters: the escape route these companies chose was owning the inference layer, not a cheaper closed API — the same decision our self-hosting versus API cost analysis breaks down.

Laya and Kev: the Jev-replica cluster goes open

Two open replicas of TypeSafe's closed Jev decision-model API shipped within days: Laya (421M, Apache 2.0, checkpoints verified on Hugging Face) and Jared Palmer's Kev family (0.5B–9B). Hacker News testers found real gaps against Jev. Why it matters: sub-35ms typed decisions are now hostable on a single consumer GPU — the discourse, and what holds up.

Watching tomorrow

Cohere Labs' claimed 218B translation MoE — the licence is still unconfirmed, and whether commercial use is permitted decides if it is deployable or a research artifact.