Agentic arms race accelerates
Multiple labs are sprinting on agentic stacks — xAI teased Grok Computer and parallel training of Grok 5 (reported 6T params, AGI‑parity claim for Q4‑2026) while Grok 4.2/4.20 betas aim at faster learning and video; Mistral launched Forge plus Mixtral 8x22B (128k context) and Small 4 (128 experts). ( ) Other pushes include MiniMax M2.7 (agentic coding) and Xiaomi’s MiMo V2 (trillion‑param MoE with browser ops) — the ecosystem is fragmenting into specialized, agent‑first models. ( )
Elon Musk has publicly described Grok 5 as a ~6‑trillion‑parameter project and has suggested it carries a nonzero chance of reaching AGI-level performance, remarks reported during investor events and on social platforms. (benzinga.com) xAI’s own timeline shows Grok 5 remains in training and the Q1 2026 window cited earlier has not produced a public release, while the company has signaled a faster cadence by running multiple Grok builds in parallel this month. (nxcode.io) (basenor.com) Grok 4.20/4.2 public beta—rolled out in mid‑February 2026—introduced a multi‑agent runtime (four cooperating agents in the beta variants) and “rapid learning” update paths intended to accelerate iterative improvements and multimodal experiments. (gigazine.net) (cometapi.com) Mistral’s new Forge platform debuted as an enterprise “build‑your‑own frontier model” service at GTC 2026, positioned to let customers pretrain and refine models on proprietary data rather than rely solely on retrieval‑augmented workflows. (mistral.ai) (techcrunch.com) Mistral’s Mistral Small 4 is a 119B‑parameter Mixture‑of‑Experts model with 128 experts (4 active per token) and a 256k context window in its documentation, and the company released it under permissive licensing alongside Forge. (mistral.ai) (docs.mistral.ai) Mixtral 8x22B—Mistral’s earlier open MoE—uses sparse activation (≈39B active parameters from ~141B total) and a long context window (≈64K tokens) and remains distributed for download and hosted use via platforms like SageMaker and Hugging Face. (docs.mistral.ai) (aws.amazon.com) MiniMax’s M2.7 launch is being billed as an “self‑evolving” agentic coding model that ran hundreds of autonomous fine‑tuning loops in internal trials and reported roughly ~30% gains on its internal benchmarks during those loops. (minimax.io) (venturebeat.com) Xiaomi published the MiMo‑V2 family (MiMo‑V2‑Pro/Omni/TTS) on March 18–19, 2026, describing a trillion‑parameter MoE flagship with ~42B active parameters, a 1,000,000‑token context mode, and native browser/tool‑use integrations exposed via MiMo Studio and the MiMo API. (mimo.xiaomi.com) (platform.xiaomimimo.com)
