Daily AI News | DeepSeek V4 Pro 0813 goes live with stronger agent capabilities
# Daily AI News | DeepSeek V4 Pro 0813 goes live with stronger agent capabilities
## Lead Story
### DeepSeek V4 Pro 0813 goes live with stronger agent capabilities
DeepSeek V4 Pro 0813 has moved out of preview into the flagship production API. In-window direct reporting and the current DeepSeek API documentation identify the 0813 version, with the update focused on agent tasks, tool use, and sustained execution.
**AI take:** The important change is production availability rather than a parameter headline; independent testing still needs to establish long-task reliability.
## More News
### SpaceXAI releases Grok 4.6 across APIs and developer platforms
SpaceXAI released Grok 4.6 with improvements for long-running agents, coding, and interactive visual work, and made it available through its API and partner developer platforms.
**AI take:** Frontier competition is moving toward sustained multi-step execution, where reliability matters more than one benchmark result.
### SpaceXAI launches Grok Bot as an always-on cloud agent
Grok Bot runs on a dedicated cloud computer, can execute multi-step work across applications, and is being offered to eligible paid Grok and Cursor users. Each bot uses a cloud desktop and browser to interact with services instead of depending on a dedicated API integration.
**AI take:** This moves xAI from answering questions toward operating accounts and apps, which also raises the stakes for credential and permission safety.
### Qwen3.8-2.4T-A95B model weights are released
Alibaba published the Qwen3.8-2.4T-A95B weights on Hugging Face, while NVIDIA released an official deployment guide for GB300 NVL72. The MoE model has 2.4 trillion total and 95 billion active parameters.
**AI take:** Open weights and a deployment path make this more than a leaderboard update, although the hardware requirement remains substantial.
### LTX-2.5 opens its weights and integrates with ComfyUI
Lightricks released LTX-2.5 weights with an emphasis on faster video generation and workflow integration, including native ComfyUI support. The official model repository makes the reviewed files directly available under its published license.
**AI take:** Open weights and ComfyUI access can move the model into real creative workflows quickly, but memory needs and license boundaries still need checking.
### Twitch defaults creators into Amazon generative-AI training
Twitch will allow parent company Amazon to use channel content for generative-AI training by default, requiring creators to turn the setting off under security and privacy.
**AI take:** The default shifts awareness and action costs to creators, so people who never inspect the setting may be included.
### Google brings sign-language-to-text AI to Pixel 11
Google DeepMind moved its multilingual SL2T model into consumer products for the first time. Gboard and Live Transcribe on Pixel 11 will initially translate American Sign Language into English.
**AI take:** This advances sign-language AI from research into daily use, although initial language coverage remains narrow.
### Liquid AI releases the edge-focused LFM2.5-VL-3B
Liquid AI released a 3.1-billion-parameter vision-language model with stronger screen understanding, grounding, multi-image input, and function calling, along with public weights and instructions.
**AI take:** Small vision models are shifting from image Q&A toward operating interfaces and tools on privacy-sensitive edge devices.
### Samsung Semiconductor uses Claude for chip development and verification
Korean and semiconductor outlets report that Samsung teams are using Anthropic Claude in SoC design and verification workflows, with some verification work said to have fallen from more than a month to a few days.
**AI take:** This shows language models entering high-bar engineering workflows, although the efficiency claims still need fuller methodology and comparison data.
### The Gemini app reaches one billion monthly active users
Google announced at Made by Google that the Gemini app has reached one billion monthly active users and called it the fastest Google product ever to reach that scale.
**AI take:** This is an adoption milestone rather than a new capability, but it affects developer, content, and enterprise ecosystem priorities.
## Rumors
### Tibo teases a “little surprise” for Codex users
OpenAI engineer Tibo Sottiaux reportedly teased a “little surprise” after Codex crossed 10 million active users. He did not say whether it concerns a usage reset, higher limits, or a larger product update.
**AI take:** The attributable teaser belongs in the report, but no specific Codex change is confirmed yet.
### Cognition reportedly seeks funding at a $40 billion valuation
AI coding company Cognition is reportedly in talks for a new round at roughly a $40 billion valuation, according to two independent outlets. No completed transaction or final terms have been announced.
**AI take:** The talks signal high expectations for AI coding, but a target valuation is not a closed funding round.
### DeepSeek Harness signals a possible general-purpose agent product
A DeepSeek Harness team public account has reportedly appeared and may signal a general-purpose agent product. DeepSeek has not published product details.
**AI take:** If confirmed, this would extend DeepSeek beyond model APIs, but today the evidence remains an organizational signal.
### Tencent Hunyuan Hy4 is reportedly coming soon
A larger Hunyuan Hy4 model will reportedly follow rapidly growing Hy3 usage. Tencent has not published its release date, specifications, pricing, or availability.
**AI take:** The cross-reported teaser is worth tracking, but evaluation must wait for an official model card and actual access.
### SK Hynix reportedly revives expansion of its Dalian NAND plant
SK Hynix is reportedly reviving or expanding construction at the Solidigm Dalian NAND facility amid AI storage demand, according to DIGITIMES and Blocks & Files. Final scale and production timing are unannounced.
**AI take:** Two industry reports clear the rumor threshold, but the project is not final until capacity, spending, and milestones are disclosed.
### Intel CEO hints at a possible return to the memory business
Intel CEO Lip-Bu Tan reportedly said he is considering new memory architectures and suggested that stacking memory on a CPU could make sense. He also said the project is not ready to be disclosed.
**AI take:** If Intel truly returns to memory or advances CPU-stacked memory, it would be a significant AI-hardware shift; today it remains a directional CEO hint.
