Journal / July 5, 2026

China open-sourced a frontier model trained without a single Nvidia chip

Meituan released LongCat-2.0 under an MIT license: 1.6 trillion parameters in a mixture-of-experts design, a native million-token context window, and 59.5% on SWE-bench Pro…

July 5, 20264 signals

Direct answer

What changed in AI on July 5, 2026?

Meituan released LongCat-2.0 under an MIT license: 1.6 trillion parameters in a mixture-of-experts design, a native million-token context window, and 59.5% on SWE-bench Pro…

Published by Bles SoftwareUpdated July 5, 20264 linked signalsHow we build the roundup

Signal digest

The sources worth opening

China open-sourced a frontier model trained without a single Nvidia chip

Meituan released LongCat-2.0 under an MIT license: 1.6 trillion parameters in a mixture-of-experts design, a native million-token context window, and 59.5% on SWE-bench Pro, trained end to end on a 50,000-card cluster of domestic Chinese ASICs. It also turned out to be Owl Alpha, the anonymous model that quietly topped OpenRouter rankings before anyone knew who built it. Export controls were supposed to make exactly this impossible. For an operator the read is that the floor keeps rising: a free, commercially usable model now does serious coding work, so any product priced on model access alone is standing on sand.

Source: Original source

Court emails show the Pentagon demanded autonomous weapons from Anthropic

Court documents revealed that Pentagon Undersecretary Emil Michael demanded Anthropic accept fully autonomous weapons use and domestic mass surveillance access as conditions of its government contracts, and that Dario Amodei refused both categories outright, citing models that are not reliable enough to pull humans out of the loop. This lands the same week OpenAI offered Washington a 5% equity stake ahead of its September IPO. The labs are now negotiating what their models will and will not do at the level of national governments, and every downstream terms-of-service inherits those choices. Know where your vendor drew its lines, because you build inside them.

Source: Original source

xAI moves to a monthly model calendar

Grok 4.5, a 1.5 trillion-parameter model, entered private beta inside SpaceX and Tesla, with Musk claiming performance close to or above Opus on internal evals. More telling than the model is the cadence: xAI says it plans monthly releases through 2026, with Grok 5 targeting 6 to 10 trillion parameters, the largest architecture anyone has discussed publicly. Meta meanwhile says its Watermelon model matches GPT-5.5 by spending an order of magnitude more compute. The treadmill is speeding up and brute force still works. Build model-agnostic, re-benchmark on a schedule you control, and stop re-platforming every time someone ships.

Source: Original source

Half a trillion in venture capital, and two labs took 43 cents of every dollar

Crunchbase puts global VC at $510 billion for the first half of 2026, with OpenAI and Anthropic alone capturing $217 billion, roughly 43% of everything invested on earth. AI as a sector absorbed an estimated 65 to 70% of all deployed capital, and Crusoe is in talks for another $3 billion for data centers. Palantir's Alex Karp called frontier pricing a wealth tax on businesses. The practical read: capital is pooling at the model layer while the application layer runs comparatively dry. If you build useful things on top of the models, your competition is less funded than the headlines make it feel.

Source: Original source

← 2026-07-04 2026-07-06 →