Aleph Alpha ships Kolibri: EU-native open weights under Apache 2.0
Aleph Alpha released Kolibri on 3 October — an English–German Mixture-of-Experts model with 78.1B total parameters and about 3.46B active per token, full weights on Hugging Face under Apache 2.0, and context up to 1M tokens. The company framed it for sovereign, on-prem deployments in regulated sectors (public administration, industrials, aerospace), with training on infrastructure in Germany and Finland and an explicit EU AI Act / GPAI Code of Practice posture. For buyers under European data-residency and audit pressure, this is a concrete open-weight alternative to US and Chinese frontier stacks — still needing two high-end GPUs at minimum, but runnable without sending prompts off-site.
What else matters
DOJ arrests CEO over alleged $300M Nvidia server smuggling to China
US prosecutors arrested Greg Lui of Earthmade Computer on charges tied to shipping more than $300 million in export-controlled GPU servers — including Nvidia A100 and H100 — to China via Malaysia and Singapore without Commerce licenses. The alleged scheme ran from late 2023 into August 2026. It is another hard enforcement signal that diversion through Southeast Asia remains a live national-security priority, not a paper rule. (DOJ; primary page blocked on fetch — details from DOJ/search reporting.)
Black Forest Labs opens FLUX 3 Image API
Black Forest Labs launched FLUX 3 Image with multi-turn edits that aim to leave untouched pixels alone, bounding-box layout, up to 10 reference images, and output up to 4K. API access is live with a launch discount through 8 October; commercial weights are available now, open weights promised in coming weeks. The race with Ideogram-class editors is now about controllable local edits, not just pretty first drafts.
Volantis raises $88M Series A for photonic AI interconnects
San Francisco startup Volantis closed an $88M Series A (co-led by Lachy Groom and Abstract Ventures) to commercialize optical interconnects that couple a GPU to large memory pools — pitched as a path past the “AI memory wall” toward multi-trillion-parameter inference, with first engines targeted for 2027. (StartupWire)
a16z leads Conway Research round for on-device Underdog
Conway Research launched Underdog, a personal assistant that runs entirely on Apple-silicon Macs with no cloud offload, and a16z is leading its first funding round (size undisclosed; Khosla and angels also cited). Capability claims are the company’s own, but the backer table is a clear bet that everyday assistant work can stay on hardware you already own. (DeAI)
Worth your time
Aleph Alpha’s Kolibri launch post and tech report — useful if you need the architecture, German data mix, and serving footprint in one place.
