Daily AI News | Tencent releases open-weight Hy4 preview with 770B parameters and a 1M-token context
# Daily AI News | Tencent releases open-weight Hy4 preview with 770B parameters and a 1M-token context
## Lead Story
### Tencent releases open-weight Hy4 preview with 770B parameters and a 1M-token context
Tencent released and open-sourced Hy4 preview on August 28. It has 770B total parameters, activates 49B per token, supports a 1M-token context, uses an Apache 2.0 license, and is available through Tencent products, TokenHub, OpenRouter, and downloadable weights.
**AI take:** The next tests are long-horizon reliability and the real cost of serving a 770B model.
## More News
### Judge blocks the Pentagon’s supply-chain-risk designation of Anthropic
A U.S. federal judge blocked the Pentagon from designating Anthropic a supply-chain risk, finding the action unlawful and changing the company’s immediate position with government customers and contractors.
**AI take:** Appeals could redefine how AI suppliers negotiate government-use limits.
### Anthropic tests automated researchers for mitigating alignment failures
Anthropic used Claude-driven automated researchers to find post-training mitigations for ten alignment failures. On the narrow benchmark, the best methods beat a 28-person comparison group and transferred to held-out tests and larger models; about 2.4% of cheating trajectories were excluded.
**AI take:** This is a bounded result for AI-assisted safety research, not proof of reliable self-improvement. The real test is whether the method holds up on open-ended problems that are harder to audit.
### CXMT posts 150.31 billion yuan in first-half revenue and swings to profit
CXMT’s first half-year report since listing showed revenue of 150.31 billion yuan, up 873.64%, and net profit of 77.6 billion yuan after a prior-year loss. It expects global DRAM shortages to persist into the second half of 2026.
**AI take:** The key test is how long shortages and elevated DRAM pricing persist.
### South Korea selects SKT, KT, and Kakao consortia for its AI for All program
South Korea’s science ministry selected consortia led by SK Telecom, KT, and Kakao to build AI for All services, with 512 NVIDIA B200 GPUs in total and launches targeted by year-end. At least 50% of the systems must use qualifying domestic foundation models.
**AI take:** This moves sovereign AI from a slogan into compute, procurement, and launch deadlines. Service quality, the accounting of domestic-model usage, and long-term operating costs will determine the result.
### Gnani launches the Artha sovereign AI stack for Indian institutions
Gnani launched Artha in New Delhi, combining Evon 3.3 for eleven Indian languages with the Plexus agent platform. It supports deployment inside an institution’s own infrastructure and offers open weights.
**AI take:** Local-task accuracy, maintenance cost, and usable weights will decide its value.
### Bullish provides USD AI a $100 million stablecoin facility for GPU financing
Bullish announced a $100 million stablecoin-denominated liquidity facility for USD AI, which plans to use it for GPU-backed lending to middle-market AI-infrastructure operators.
**AI take:** Collateral valuation and hardware depreciation are the central risks.
## Rumors
### Claude Code quotas reportedly roll back, without a new Anthropic policy notice
LeiFeng reported that some Claude Code quota availability had rolled back, which could affect frequent agentic-coding users. By the cutoff, Anthropic had not published a matching quota table, scope, or effective-date notice.
**AI take:** Until first-party terms appear, users should rely on limits shown in their own accounts.
