2026-09-15 · TuesdayPublished daily at 08:00
AI Daily
- 1Xiaohongshu’s AllSpark open-sources the Search Agent model Iris, with the 35B and 397B versions achieving leading results among models of comparable sizeWeChat Official Account: Xiaohongshu Technology
- 2
- 3
01Model Releases/Updates
Xiaohongshu’s AllSpark open-sources the Search Agent model Iris, with the 35B and 397B versions achieving leading results among models of comparable size
Xiaohongshu’s AllSpark team releases and open-sources the Search Agent model Iris. The weights and evaluation code are now public, while the data and training recipe will be released in stages.
Original textSiliconFlow launches open-source model Hy4 preview, with 770B total parameters and a 1M context window
SiliconFlow announced that the open-source model Hy4 preview is now available on its platform. The model has 770B total parameters, activates 49B per token, supports a 1M context window, and is released under the Apache 2.0 license. It is designed for coding, analysis, research, and complex real-world work. Users can connect it to existing tools such as Claude Code, Codex, and Cursor. The image shows pricing per 1M tokens: $0.834 for input, $2.501 for output, and $0.042 for cached tokens.
Original text02Product Releases/Updates
Apple launches next-generation Apple Intelligence, with Siri AI officially arriving in beta
Apple released the next generation of Apple Intelligence, introducing a comprehensively rebuilt Siri AI with support for understanding personal context, screen awareness, system-level app operations, and cross-device conversations. Beginning today, it will roll out as an English beta with the 2027 system updates, expanding next month to French, Japanese, Korean, Portuguese, and Spanish.
Original text03Industry News
Anthropic plans Nasdaq listing after posting profits for a second consecutive quarter, targeting a $2 trillion valuation
Anthropic told investors it would achieve profitability for the second consecutive quarter, but the claim is based on adjusted metrics that exclude costs such as stock-based compensation; according to the Financial Times, its gross margin exceeds 80%, before accounting for revenue shares with partners such as Amazon and model training costs.
Original text04Paper research
DeepSeek-V4.1-Flash (Max) ranks No. 3 among open-source models in Agent Arena
DeepSeek-V4.1-Flash (Max) entered third place among open-source models in Agent Arena, with a net improvement of +4.87% and a median cost per task of $0.07, reshaping the Pareto frontier. Its cost is 68% lower than that of second-place Hy4 preview, while its score is only 0.09 percentage points lower. It ranks 12th overall and fourth in the Confirmed Success signal ranking (+13.75%).
Original text05Tips and perspectives
Is the tech giants’ verbal agreement to slow AI development a safety consensus or a cartel?
Sam Altman, Dario Amodei, Demis Hassabis, and Elon Musk broadly agreed over the weekend to slow AI development, proposing third-party audits, regulation of domestic laboratories, and a global slowdown agreement. Critics called it a “cartel” aimed at suppressing competitors and the open-source movement.
Original textGPT-5.6 Luna vs. GPT-6 Astra: Is a $1.20 model good enough for code review?
Entelligence compared the code-review capabilities of GPT-5.6 Luna and GPT-6 Astra on 50 public benchmark PRs using the same prompts. Luna found 69 verified bugs, compared with Astra's 92, at a total cost of $0.20 versus $5.66 and an accuracy of 74% versus 96%.
Original textTomer Tunguz analyzes the five camps behind Amodei’s proposal to slow frontier AI and the challenges of regulating compute
Tomer Tunguz analyzes Dario Amodei’s call to slow frontier AI development, identifying five camps focused on interpretability, labor, economics, geopolitics, and regulatory capture, and points out that none of them specifies a concrete pace.
Original textHow Anthropic reengineered its test impact analysis service to handle the CI pressure of agentic coding
Anthropic engineers deliver 8 times as much code per quarter as the 2021-2025 average, with Claude writing 80% of it. CI tasks increased 25-fold within six months.
Original text