AI News Sep 21, 2026
By Frontier Editorial •
Key Takeaways
- DeepSeek CEO calls training on Huawei chips one of its biggest bets: DeepSeek CEO Liang Wenfeng told investors that training its models on Huawei chips is one of the company's bigg…
- Unitree releases partial weights for 6-billion-parameter UnifoLM model: Unitree has released partial weights and data for UnifoLM-WLA-1.0, its 6-billion-parameter world-language-ac…
- MiniMax open-sources its terminal-based MiniMax Code agent: MiniMax has open-sourced MiniMax Code, a terminal-based coding agent released under the MIT license. The agent can under…
- 1. Open your Claude Code configuration settings.
What are the top AI breakthroughs?
This Sep 21, 2026 covers 9 curated AI news items spanning technology, research, and product developments. Alibaba releases 7B-parameter Qwen-Image-2.1 for consumer GPUs: Alibaba's Qwen team has released Qwen-Image-2.1, an open...
Alibaba releases 7B-parameter Qwen-Image-2.1 for consumer GPUs: Alibaba's Qwen team has released Qwe…
Alibaba releases 7B-parameter Qwen-Image-2.1 for consumer GPUs: Alibaba's Qwen team has released Qwen-Image-2.1, an open-weight 7-billion-parameter model capable of generating and editing images on consumer GPUs. The model supports transparency and up to ten reference images, but its research license restricts commercial use without a separate Qwen license.
StepFun releases Step 5 Preview open-source model with 27B active parameters: StepFun has released S…
StepFun releases Step 5 Preview open-source model with 27B active parameters: StepFun has released Step 5 Preview, an open-source model featuring 27 billion active parameters. The model ranks among the top two open-source models in recent evaluations, demonstrating strong performance despite its compact active parameter size.
Amazon blocks Meta's Muse AI assistant from shopping on its platform: Amazon has blocked Meta's new …
Amazon blocks Meta's Muse AI assistant from shopping on its platform: Amazon has blocked Meta's new Muse personal AI assistant from accessing Amazon.com to shop on behalf of users. The e-commerce giant cited terms of service violations, security risks, and a lack of merchant consent for the decision.
Researchers introduce two-step test-time training protocol for VARC models: Researchers have introdu…
Researchers introduce two-step test-time training protocol for VARC models: Researchers have introduced a novel two-step test-time training (TTT) protocol to improve rule induction in Vision ARC (VARC) models. The method first finetunes only the task embedding representing the transformation rule, and then freezes it to finetune the backbone.
LogicTrack framework uses formal logic solvers to audit LLM reasoning steps: Researchers have propos…
LogicTrack framework uses formal logic solvers to audit LLM reasoning steps: Researchers have proposed LogicTrack, a neuro-symbolic framework designed to verify the logical validity of intermediate reasoning steps in large language models. The system auto-formalizes each reasoning step into symbolic representations and verifies them using automated theorem provers.
DeepSeek CEO calls training on Huawei chips one of its biggest bets: DeepSeek CEO Liang Wenfeng told…
DeepSeek CEO calls training on Huawei chips one of its biggest bets: DeepSeek CEO Liang Wenfeng told investors that training its models on Huawei chips is one of the company's biggest bets. Huawei is reportedly scheduled to deliver its next-generation training chips to DeepSeek in the fourth quarter of 2026 or the first quarter of 2027.
Unitree releases partial weights for 6-billion-parameter UnifoLM model: Unitree has released partial…
Unitree releases partial weights for 6-billion-parameter UnifoLM model: Unitree has released partial weights and data for UnifoLM-WLA-1.0, its 6-billion-parameter world-language-action model for humanoid robots. The model covers 64 tasks, though post-training code and full datasets are still planned for future release.
MiniMax open-sources its terminal-based MiniMax Code agent: MiniMax has open-sourced MiniMax Code, a…
MiniMax open-sources its terminal-based MiniMax Code agent: MiniMax has open-sourced MiniMax Code, a terminal-based coding agent released under the MIT license. The agent can understand projects, modify files, and run tests via a command-line interface, supporting both an interactive terminal UI and a headless execution mode.
RBS-Attention training-free method optimizes long-context LLM prefill: To address the prefill bottle…
RBS-Attention training-free method optimizes long-context LLM prefill: To address the prefill bottleneck in long-context large language model inference, researchers proposed RBS-Attention, a training-free sparse-prefill method. The approach uses a centroid base branch to capture average relevance and a rescue branch to identify blocks at risk of underestimation due to mean dilution.
What are the latest AI investment signals?
Latest AI investment signals: 3 funding rounds, 0 market updates, and 0 M&A transactions.
Primary Market – Funding Rounds
| Company | Amount | Round |
|---|---|---|
| Techmeme | — | Nvidia |
| Techmeme | — | Humanoid Robots |
| Metris Energy | $5,000,000 | seed |
Secondary Market – Market Updates
No secondary market data.
M&A – Mergers & Acquisitions
No M&A data.
What are practical AI tips this week?
3 practical AI tips curated from Reddit communities and expert blogs. Disabling Claude Code prompt suggestions saves up to 10% of usage limits...
Claude Code
Disabling Claude Code prompt suggestions saves up to 10% of usage limits
1. Open your Claude Code configuration settings. 2. Locate the setting for 'Prompt Suggestions'. 3. Turn off or disable the Prompt Suggestions feature to stop automatic context cache reads. 4. Avoid compacting your context immediately after the cache expires to further conserve limits.
AI Image Generation
One-line prompt fix prevents AI from altering product details in ads
1. Upload your original product photo to the AI image generation tool as a reference. 2. Write your prompt describing the desired background, lighting, and mood. 3. Add an explicit rule stating that the reference photo is the single source of truth for the product's physical state. 4. Instruct the AI not to alter, open, or modify any physical details of the product packaging.
Prompt Engineering
Two-step prompting method prevents AI from generating generic slide decks
1. Prompt your AI to act as a skeptic and generate a single core argument for your deck. 2. Ask the AI to outline one message per slide that builds this argument, excluding any design talk. 3. Review and refine this structural spine yourself. 4. Input the finalized structure into an AI presentation generator like Gamma to lay out the slides.