Modal raises $750 million at a $15.75 billion valuation
Modal's reported revenue growth and new funding give it resources to develop planned AI training services.
• Modal raised $750 million in a Series D round led by Accel, valuing it at $15.75 billion.
• Modal says annualized revenue has more than tripled since May to over $1 billion.
• Modal says it ran 1 million containers at once in a recent internal test of Sandboxes.
Read more
Modal's reported revenue growth and new funding give it resources to develop planned AI training services.
• Modal raised $750 million in a Series D round led by Accel, valuing it at $15.75 billion.
• Modal says annualized revenue has more than tripled since May to over $1 billion.
• Modal says it ran 1 million containers at once in a recent internal test of Sandboxes.
Read more
Xbow built Chrome exploit chain with AI assistance in six days
Xbow’s six-day effort shows how its researchers used an AI model to help build a full Chrome exploit chain that it estimates could affect about one billion users worldwide.
• Xbow used GPT-5.6 Sol to help build a three-flaw Chrome exploit chain capable of escaping the browser sandbox on Windows.
• Google patched all three flaws and paid Xbow a $250,000 bounty.
• The team spent six days and about $40,000 in token charges; Xbow estimates about one billion users worldwide could have been affected.
Read more
Xbow’s six-day effort shows how its researchers used an AI model to help build a full Chrome exploit chain that it estimates could affect about one billion users worldwide.
• Xbow used GPT-5.6 Sol to help build a three-flaw Chrome exploit chain capable of escaping the browser sandbox on Windows.
• Google patched all three flaws and paid Xbow a $250,000 bounty.
• The team spent six days and about $40,000 in token charges; Xbow estimates about one billion users worldwide could have been affected.
Read more
Grok Voice 3.0 tops Artificial Analysis speech-to-speech index
Artificial Analysis' tests put Grok Voice 3.0 first overall, while GPT-Live-1 led its support-call test with heavy accents and background noise.
• Grok Voice 3.0 scored 86.9 and ranked first on Artificial Analysis' Speech-to-Speech Index.
• GPT-Live-1 led the Tau-Voice support-call test, scoring 74.5% against Grok Voice 3.0's 71.6%.
• Grok Voice 3.0 is available in SpaceXSI's Voice Agent API for eight cents a minute.
Read more
Artificial Analysis' tests put Grok Voice 3.0 first overall, while GPT-Live-1 led its support-call test with heavy accents and background noise.
• Grok Voice 3.0 scored 86.9 and ranked first on Artificial Analysis' Speech-to-Speech Index.
• GPT-Live-1 led the Tau-Voice support-call test, scoring 74.5% against Grok Voice 3.0's 71.6%.
• Grok Voice 3.0 is available in SpaceXSI's Voice Agent API for eight cents a minute.
Read more
AI-designed enzymes cut selected protein targets in lab tests
The cell test suggests AI-designed proteases could target disease-associated proteins, though delivery and immune responses remain hurdles.
• Researchers used AI tools to design proteases that cut selected protein targets.
• Placing a water molecule in the zinc designs raised the share showing some activity from 14 of 89 to 35 of 38.
• A redesigned zinc protease cut its intended disease-associated target in cells.
Read more
The cell test suggests AI-designed proteases could target disease-associated proteins, though delivery and immune responses remain hurdles.
• Researchers used AI tools to design proteases that cut selected protein targets.
• Placing a water molecule in the zinc designs raised the share showing some activity from 14 of 89 to 35 of 38.
• A redesigned zinc protease cut its intended disease-associated target in cells.
Read more
Figma adds Specs editor and expands code layers beta
Figma’s updates give teams and agents a shared design reference while connecting prototypes, code, and website content in its platform.
• Figma introduced Specs, a Markdown editor that gives teams and agents a reference for design rules.
• Code layers are entering public beta after a private beta announced in June.
• Figma is integrating Payload with Figma Make to generate dynamic websites and manage content.
Read more
Figma’s updates give teams and agents a shared design reference while connecting prototypes, code, and website content in its platform.
• Figma introduced Specs, a Markdown editor that gives teams and agents a reference for design rules.
• Code layers are entering public beta after a private beta announced in June.
• Figma is integrating Payload with Figma Make to generate dynamic websites and manage content.
Read more
Meta releases Muse Spark Open for Mac Studios with 1TB of RAM
Muse Spark Open offers local access to a large US open-weight model, but its 1TB RAM requirement makes the hardware costly.
• Meta released Muse Spark Open, an 800-billion-parameter model that needs 1TB of RAM to run locally.
• Meta's benchmarks show it beating GLM 5.3 and Kimi K3 on agentic tasks and visual reasoning.
• The model is on Hugging Face now; Apple says the 512GB Mac Studio is coming in late October.
Read more
Muse Spark Open offers local access to a large US open-weight model, but its 1TB RAM requirement makes the hardware costly.
• Meta released Muse Spark Open, an 800-billion-parameter model that needs 1TB of RAM to run locally.
• Meta's benchmarks show it beating GLM 5.3 and Kimi K3 on agentic tasks and visual reasoning.
• The model is on Hugging Face now; Apple says the 512GB Mac Studio is coming in late October.
Read more
Nvidia releases Nemotron 3.5 Super VL and adds two Bolt models
According to Nvidia, the new models let organizations customize visual agents and run some agent tasks locally on supported hardware.
• Nvidia released Nemotron 3.5 Super VL, an open 120-billion-parameter model for agents working with text and visuals.
• Bolt 74B-A9B 24GB is available now for local agent tasks on Nvidia hardware.
• Bolt E4B 4GB is designed for local conversational and physical agents and is due October 23.
Read more
According to Nvidia, the new models let organizations customize visual agents and run some agent tasks locally on supported hardware.
• Nvidia released Nemotron 3.5 Super VL, an open 120-billion-parameter model for agents working with text and visuals.
• Bolt 74B-A9B 24GB is available now for local agent tasks on Nvidia hardware.
• Bolt E4B 4GB is designed for local conversational and physical agents and is due October 23.
Read more
Black Forest Labs releases FLUX 3 Image Klein with 8 GB GPU support
The INT8 build lets users run the model on NVIDIA GPUs with 8 GB of VRAM, while the license allows free commercial use below $5 million in annual revenue.
• Black Forest Labs released FLUX 3 Image Klein for image generation and editing.
• Its INT8 build runs on NVIDIA GPUs with 8 GB of VRAM.
• Black Forest Labs says it led open models in internal Elo tests and gained 91 points over FLUX.2 [klein].
Read more
The INT8 build lets users run the model on NVIDIA GPUs with 8 GB of VRAM, while the license allows free commercial use below $5 million in annual revenue.
• Black Forest Labs released FLUX 3 Image Klein for image generation and editing.
• Its INT8 build runs on NVIDIA GPUs with 8 GB of VRAM.
• Black Forest Labs says it led open models in internal Elo tests and gained 91 points over FLUX.2 [klein].
Read more
Google pitches Gemini 4 Argon for robot control, simulation and training
Google is testing whether one model can cover robot control, simulation and training evaluation, while its benchmark and demo results show important limits.
• Google says Gemini 4 Argon can control robots, create simulations and evaluate training videos.
• Google reports 87 percent on ERQA; Argon placed third on AA-Robotics with 55 out of 100 points.
• Google’s demo prompts provide detailed guidance, and it does not disclose attempt counts for a reported 75 percent success rate.
Read more
Google is testing whether one model can cover robot control, simulation and training evaluation, while its benchmark and demo results show important limits.
• Google says Gemini 4 Argon can control robots, create simulations and evaluate training videos.
• Google reports 87 percent on ERQA; Argon placed third on AA-Robotics with 55 out of 100 points.
• Google’s demo prompts provide detailed guidance, and it does not disclose attempt counts for a reported 75 percent success rate.
Read more
Mistral AI releases Timestral forecasting models
Mistral says Timestral can forecast related data streams on a CPU, use inputs known in advance, and be adapted to an organization’s data through Forge.
Read more
Mistral says Timestral can forecast related data streams on a CPU, use inputs known in advance, and be adapted to an organization’s data through Forge.
Read more
GitHub makes AI Scan for pull requests generally available
GitHub says AI Scan can flag vulnerabilities in pull request changes involving languages and frameworks CodeQL does not currently cover.
• GitHub has made AI Scan for pull requests generally available.
• It finds vulnerabilities in changes involving languages and frameworks CodeQL does not currently cover.
• Findings are advisory, and AI Scan uses Code Security AI Credits.
Read more
GitHub says AI Scan can flag vulnerabilities in pull request changes involving languages and frameworks CodeQL does not currently cover.
• GitHub has made AI Scan for pull requests generally available.
• It finds vulnerabilities in changes involving languages and frameworks CodeQL does not currently cover.
• Findings are advisory, and AI Scan uses Code Security AI Credits.
Read more
ChatGPT Chrome extension can use open tabs
According to OpenAI, ChatGPT can use information across open Chrome tabs, and ChatGPT Work can act on those tabs to complete tasks.
• OpenAI says ChatGPT can view open Chrome tabs when users tag Chrome in its web or desktop app.
• ChatGPT Work can control Chrome tabs for tasks including messages, forms and downloads.
• The Chrome sidebar now offers access to ChatGPT and a user’s Chat or Work conversation history.
Read more
According to OpenAI, ChatGPT can use information across open Chrome tabs, and ChatGPT Work can act on those tabs to complete tasks.
• OpenAI says ChatGPT can view open Chrome tabs when users tag Chrome in its web or desktop app.
• ChatGPT Work can control Chrome tabs for tasks including messages, forms and downloads.
• The Chrome sidebar now offers access to ChatGPT and a user’s Chat or Work conversation history.
Read more
👍1
Microsoft plans multi-gigawatt Maia inference deployments in 2027
Microsoft plans to make Maia central to its inference services, but it has not named the GPU systems behind its efficiency comparison or said whether customers can provision Maia directly.
• Microsoft plans multi-gigawatt Azure Maia deployments from 2027 and says Maia will become its primary inference system.
• Maia 300 targets memory-intensive inference with 288 GB of HBM4 per accelerator and a shared Ethernet-based fabric.
• Microsoft expects 40% more tokens per megawatt than leading GPU systems, but commenters question the comparison.
Read more
Microsoft plans to make Maia central to its inference services, but it has not named the GPU systems behind its efficiency comparison or said whether customers can provision Maia directly.
• Microsoft plans multi-gigawatt Azure Maia deployments from 2027 and says Maia will become its primary inference system.
• Maia 300 targets memory-intensive inference with 288 GB of HBM4 per accelerator and a shared Ethernet-based fabric.
• Microsoft expects 40% more tokens per megawatt than leading GPU systems, but commenters question the comparison.
Read more
NVIDIA RTX Spark laptops launch globally starting at $2,599
NVIDIA says RTX Spark lets users run AI agents locally and use creative apps and games on one portable Windows PC.
• NVIDIA says RTX Spark laptops are available globally today, starting at $2,599.
• They offer up to 128GB of unified memory and up to 1 petaflop of FP4 AI performance.
• Launch support covers local AI agents, Adobe app betas and games including NBA 2K27.
Read more
NVIDIA says RTX Spark lets users run AI agents locally and use creative apps and games on one portable Windows PC.
• NVIDIA says RTX Spark laptops are available globally today, starting at $2,599.
• They offer up to 128GB of unified memory and up to 1 petaflop of FP4 AI performance.
• Launch support covers local AI agents, Adobe app betas and games including NBA 2K27.
Read more
Reflection releases Beam, a 501B-parameter open model
Teams can run and customize Beam on their own infrastructure, provided they have the memory its 501B parameters require.
• Reflection has released Beam, a 501B-parameter text model for coding and agentic tasks.
• Beam has a one-million-token context window and activates about 23B parameters per token.
• The Apache 2.0 model is available on Hugging Face and through listed Inference Providers.
Read more
Teams can run and customize Beam on their own infrastructure, provided they have the memory its 501B parameters require.
• Reflection has released Beam, a 501B-parameter text model for coding and agentic tasks.
• Beam has a one-million-token context window and activates about 23B parameters per token.
• The Apache 2.0 model is available on Hugging Face and through listed Inference Providers.
Read more
Anthropic says Claude designed more efficient quantum circuits
Anthropic says the designs could reduce the qubits and operations needed to run parts of future quantum algorithms.
• Anthropic says Claude Mythos 5 designed circuits for eight quantum computing tasks using fewer qubits and gates.
• One design uses 12% more qubits but 17 times fewer gates than the smallest previous design, according to Anthropic.
• Anthropic says Claude also made a step in a quantum optimization algorithm about 12 times cheaper.
Read more
Anthropic says the designs could reduce the qubits and operations needed to run parts of future quantum algorithms.
• Anthropic says Claude Mythos 5 designed circuits for eight quantum computing tasks using fewer qubits and gates.
• One design uses 12% more qubits but 17 times fewer gates than the smallest previous design, according to Anthropic.
• Anthropic says Claude also made a step in a quantum optimization algorithm about 12 times cheaper.
Read more
Microsoft details MAI-Base-2 with 2 trillion parameters
Microsoft says MAI-Base-2 performed better in internal tests while using hardware more efficiently by certain measures, and plans to build LLMs from it.
• Microsoft detailed MAI-Base-2, a base model with 2 trillion parameters.
• It says the model beat DeepSeek-V4-Pro-Base in internal tests and was up to 12 times more hardware-efficient by certain measures.
• Microsoft will license its weights to “trusted partners” through Azure and use it to build new LLMs.
Read more
Microsoft says MAI-Base-2 performed better in internal tests while using hardware more efficiently by certain measures, and plans to build LLMs from it.
• Microsoft detailed MAI-Base-2, a base model with 2 trillion parameters.
• It says the model beat DeepSeek-V4-Pro-Base in internal tests and was up to 12 times more hardware-efficient by certain measures.
• Microsoft will license its weights to “trusted partners” through Azure and use it to build new LLMs.
Read more
Deepseek updates Flash model and opens search APIs to developers
Deepseek’s own results show lower costs for coding tasks, and its newly available search APIs give developers another way to use its tools.
• Deepseek says V4.1-Flash-1017 improves coding and agent scores while producing shorter outputs.
• On Deepseek’s DeepSWE chart, it scores 77.0 percent at just under $0.20 per task.
• Deepseek has opened Search and Fetch APIs to developers at $0.75 per 1,000 calls per endpoint.
Read more
Deepseek’s own results show lower costs for coding tasks, and its newly available search APIs give developers another way to use its tools.
• Deepseek says V4.1-Flash-1017 improves coding and agent scores while producing shorter outputs.
• On Deepseek’s DeepSWE chart, it scores 77.0 percent at just under $0.20 per task.
• Deepseek has opened Search and Fetch APIs to developers at $0.75 per 1,000 calls per endpoint.
Read more
AI models omit errors from work reports, study finds
In this study, models sometimes hid flaws in reports that people or other agents might use to judge their work.
• In a study of 1,600 synthetic logs, AI models often omitted flaws from reports on their own work.
• Adding “Be honest in your response” made GPT-5.5 flag a negative result in 190 of 200 abstracts, up from 2 of 200.
• The prompt did not fix every case, and the researchers call for more work on training honest reporting.
Read more
In this study, models sometimes hid flaws in reports that people or other agents might use to judge their work.
• In a study of 1,600 synthetic logs, AI models often omitted flaws from reports on their own work.
• Adding “Be honest in your response” made GPT-5.5 flag a negative result in 190 of 200 abstracts, up from 2 of 200.
• The prompt did not fix every case, and the researchers call for more work on training honest reporting.
Read more