DeAI Daily Brief — 4 October 2026

Today in DeAI: Aleph Alpha releases the fully open Kolibri-1, Prime Intellect enters open-model serving on GB200 hardware, a Nesa Chain exploit mints $55M in NES, IBM ships self-hosted Bob, and Google puts confidential GPUs into mainstream cloud SKUs.

Aleph Alpha open-sources Kolibri-1: 78B total, 3.46B active, Apache 2.0

The Heidelberg-based lab released full weights on Hugging Face on October 3 — a bilingual English-German mixture-of-experts model with a 1M-token context, trained on 20T tokens on 768 NVIDIA B200s, with Aleph Alpha citing infrastructure in Germany and Finland. Benchmarks like 96.9 on AIME 2025 are the vendor's own runs, and the comparison field predates the current open-weight leaders. Why it matters: regulated European buyers get a permissively licensed, fully downloadable model sized for a single H200-class node. (Aleph Alpha, model card) — our full coverage

Prime Intellect launches Prime Inference for frontier open models

The RL-environments company now serves an OpenAI-compatible API for frontier open-weight checkpoints on GB200 NVL72 clusters, with serverless and reserved capacity and Z.ai's GLM-5.3 live at launch. Its ~100-token-per-second end-to-end figures and stress-test claims are self-reported, and reserved-capacity pricing is not yet public. Why it matters: another venue serves the newest open weights on top-tier hardware on day one, widening the field builders price against. (Prime Intellect) — how the field compares

A Nesa Chain exploit mints and dumps roughly $55M in NES

On-chain reporting shows one wallet minted about $55M of the decentralized-inference network's token and sold into DEX liquidity; the token fell around 40%, and the team is directing users to revised canonical token contracts on Solana, Base, and Arbitrum. Figures are third-party on-chain readouts pending a full post-mortem, and the team's "widely used attack vector" characterization is its own. Why it matters: in token-settled inference networks, the settlement layer is part of the trust surface. (BigGo Finance) — decentralized inference, the landscape

IBM makes Bob self-hosted GA for on-prem and air-gapped fleets

IBM's October 1 release puts its agentic software-development platform into general availability for on-premises, sovereign-cloud, and air-gapped deployment, with supported self-hosted models — NVIDIA Nemotron and Poolside Laguna are named — or hybrid connections to external model services. Which configurations give full isolation depends on that supported-model list, and the sovereignty framing is IBM's positioning. Why it matters: "bring the AI to the data" is now a purchasable default rather than a bespoke project. (IBM)

Google puts confidential GPUs into mainstream cloud SKUs

Confidential G4 VMs — NVIDIA Blackwell-class GPUs inside confidential computing — and Intel TDX-based C4 Confidential VMs entered global preview per the Confidential Computing Consortium's roundup of Google Cloud Next announcements. Preview status means no GA SLAs, and NVIDIA's protection framing is vendor marketing until attestation tooling is independently reviewed. Why it matters: hardware-attested inference is becoming a checkbox on hosted platforms rather than a specialty project. (Confidential Computing Consortium) — our TEE coverage

Watching tomorrow

The Microsoft-NVIDIA RTX Spark event on October 7 is the one confirmed date on the board; Meta, Mistral, Moonshot, and Qwen release dates clustering around it are cadence projections, not announcements.