EmbeddingGemma 2 supports search across text, code, images, video and audio
The shared embedding space lets developers compare different media types for retrieval, while shorter vectors can reduce storage use at a cost to quality.
• EmbeddingGemma 2 represents text, code, images, video and audio in one 768-dimensional space.
• Its text and code configuration has 270 million parameters; full multimodal support has 740 million.
• At 128 dimensions, it retains about 90% of text retrieval quality and around 75% for image, video and speech retrieval.
Read more
The shared embedding space lets developers compare different media types for retrieval, while shorter vectors can reduce storage use at a cost to quality.
• EmbeddingGemma 2 represents text, code, images, video and audio in one 768-dimensional space.
• Its text and code configuration has 270 million parameters; full multimodal support has 740 million.
• At 128 dimensions, it retains about 90% of text retrieval quality and around 75% for image, video and speech retrieval.
Read more
Microsoft makes Foundry browser automation generally available
According to Microsoft, Foundry agents can complete website-based workflows, including those involving private applications, with human oversight.
• Microsoft says its Browser Automation Tool is generally available in Foundry for agents that work through websites.
• Agents can use managed browser sessions for public sites and private applications connected through an Azure virtual network.
• Live View and Take control let people watch sessions and temporarily operate the browser.
Read more
According to Microsoft, Foundry agents can complete website-based workflows, including those involving private applications, with human oversight.
• Microsoft says its Browser Automation Tool is generally available in Foundry for agents that work through websites.
• Agents can use managed browser sessions for public sites and private applications connected through an Azure virtual network.
• Live View and Take control let people watch sessions and temporarily operate the browser.
Read more
Tenstorrent raises $1.5 billion at a $6.2 billion valuation
The funding gives Tenstorrent capital to scale its AI chips as hardware startups seek to challenge Nvidia’s dominance.
• Tenstorrent raised $1.5 billion at a $6.2 billion valuation, more than double its late 2024 valuation.
• Trace Capital and Innovation X led the round after reported acquisition interest from Intel and Qualcomm.
• The company plans to scale chips for serving open-weight AI models and video generation.
Read more
The funding gives Tenstorrent capital to scale its AI chips as hardware startups seek to challenge Nvidia’s dominance.
• Tenstorrent raised $1.5 billion at a $6.2 billion valuation, more than double its late 2024 valuation.
• Trace Capital and Innovation X led the round after reported acquisition interest from Intel and Qualcomm.
• The company plans to scale chips for serving open-weight AI models and video generation.
Read more
Hacker used AI agents to break into more than 100 websites
The breaches show how an attacker used AI agents to carry out attacks across more than 100 websites, according to Norforge.
• Norforge found a hacker used mostly autonomous AI agents to break into more than 100 websites.
• Reuters identified John Deere, SharkNinja and Cummins among the targets.
• Norforge found more than 20 exposed Cairn instances it considers likely malicious and is working to get them taken down.
Read more
The breaches show how an attacker used AI agents to carry out attacks across more than 100 websites, according to Norforge.
• Norforge found a hacker used mostly autonomous AI agents to break into more than 100 websites.
• Reuters identified John Deere, SharkNinja and Cummins among the targets.
• Norforge found more than 20 exposed Cairn instances it considers likely malicious and is working to get them taken down.
Read more
Meta says Prometheus supercluster now spans hundreds of thousands of GPUs
Meta says Prometheus already trains its Muse models and is expanding toward roughly 1 million GPUs for future models.
• Meta says Prometheus now connects hundreds of thousands of GPUs in Ohio and is planned to reach roughly 1 million.
• Its first completed phase trained Muse models, including Muse Wonder, according to the company.
• Meta reports a 99% drop in training interruptions from a peak reached in mid-2025.
Read more
Meta says Prometheus already trains its Muse models and is expanding toward roughly 1 million GPUs for future models.
• Meta says Prometheus now connects hundreds of thousands of GPUs in Ohio and is planned to reach roughly 1 million.
• Its first completed phase trained Muse models, including Muse Wonder, according to the company.
• Meta reports a 99% drop in training interruptions from a peak reached in mid-2025.
Read more
Nvidia releases open models for enterprise agents and local PCs
Nvidia’s new models give organizations an option to run agents on their own infrastructure and let supported PCs handle tasks locally.
• Nvidia launched the 120-billion-parameter Nemotron 3.5 Super VL for enterprise agents.
• A 74-billion-parameter Bolt model is available for local agent tasks on PCs and workstations.
• Nvidia says the smaller Bolt E4B for conversational agents is coming later this month.
Read more
Nvidia’s new models give organizations an option to run agents on their own infrastructure and let supported PCs handle tasks locally.
• Nvidia launched the 120-billion-parameter Nemotron 3.5 Super VL for enterprise agents.
• A 74-billion-parameter Bolt model is available for local agent tasks on PCs and workstations.
• Nvidia says the smaller Bolt E4B for conversational agents is coming later this month.
Read more
Amazon says TorchNeuron runs PyTorch code on Trainium without a rewrite
Amazon says PyTorch developers can run existing code on its Trainium AI chips without rewriting it.
• Amazon says TorchNeuron runs existing PyTorch code on Trainium without a rewrite.
• The company says startups and enterprises are already building with it across several fields.
• Amazon plans to open-source TorchNeuron at the PyTorch Conference on October 20.
Read more
Amazon says PyTorch developers can run existing code on its Trainium AI chips without rewriting it.
• Amazon says TorchNeuron runs existing PyTorch code on Trainium without a rewrite.
• The company says startups and enterprises are already building with it across several fields.
• Amazon plans to open-source TorchNeuron at the PyTorch Conference on October 20.
Read more
OpenCode breach exposes API keys for some Console users
The exposed keys put affected users at risk of unauthorized API use, which multiple users reported alongside charges and project suspensions.
• OpenCode says a breach exposed stored API keys for 1.4% of Console users who linked external AI providers.
• Multiple users reported unauthorized charges; some said Google suspended their projects.
• OpenCode says it is working with Google, OpenAI and Anthropic to address the incident.
Read more
The exposed keys put affected users at risk of unauthorized API use, which multiple users reported alongside charges and project suspensions.
• OpenCode says a breach exposed stored API keys for 1.4% of Console users who linked external AI providers.
• Multiple users reported unauthorized charges; some said Google suspended their projects.
• OpenCode says it is working with Google, OpenAI and Anthropic to address the incident.
Read more
Kling AI releases version 4.0 with longer clips and a Flash model
Kling AI 4.0 gives users longer clips and more reference controls, while its Flash variant offers faster generation with lower output limits.
• Kling AI has released version 4.0 with longer clips, more scene controls and stereo audio.
• Users can generate up to 30 seconds in one pass and guide scenes with up to ten keyframe images.
• Kling AI says its faster, lower-cost Flash model supports clips up to 20 seconds at 720p.
Read more
Kling AI 4.0 gives users longer clips and more reference controls, while its Flash variant offers faster generation with lower output limits.
• Kling AI has released version 4.0 with longer clips, more scene controls and stereo audio.
• Users can generate up to 30 seconds in one pass and guide scenes with up to ten keyframe images.
• Kling AI says its faster, lower-cost Flash model supports clips up to 20 seconds at 720p.
Read more
MiniMax Releases M3.1 Flash Reasoning Model With Lower Costs
The release highlights ongoing price competition among Chinese AI developers as companies push reasoning performance higher while reducing task costs.
• MiniMax released the M3.1 Flash reasoning model with a score of 42 on the Intelligence Index at less than half the cost of its predecessor.
• The model scores 48 percent on Humanity's Last Exam and 56 percent on factual accuracy in the AA-Omniscience test.
• MiniMax priced the API at $0.15 per million input tokens and plans an open-weights release on October 23.
Read more
The release highlights ongoing price competition among Chinese AI developers as companies push reasoning performance higher while reducing task costs.
• MiniMax released the M3.1 Flash reasoning model with a score of 42 on the Intelligence Index at less than half the cost of its predecessor.
• The model scores 48 percent on Humanity's Last Exam and 56 percent on factual accuracy in the AA-Omniscience test.
• MiniMax priced the API at $0.15 per million input tokens and plans an open-weights release on October 23.
Read more
MiniMax launches M3.1 Flash, claims 70% lower AI agent costs
MiniMax’s claimed cost and speed gains target AI agents, which may need dozens of model calls to finish a single job.
• MiniMax says M3.1 Flash cuts agent task costs by 70% versus its previous model.
• The company reports nearly three times faster inference and more than 200 output tokens per second at everyday peak.
• Flash is available through MiniMax’s API, M Plan subscription and MiniMax Code app.
Read more
MiniMax’s claimed cost and speed gains target AI agents, which may need dozens of model calls to finish a single job.
• MiniMax says M3.1 Flash cuts agent task costs by 70% versus its previous model.
• The company reports nearly three times faster inference and more than 200 output tokens per second at everyday peak.
• Flash is available through MiniMax’s API, M Plan subscription and MiniMax Code app.
Read more
Nvidia and Dell join $200 million funding round for doubleAI
According to doubleAI, Maxwell can make software more efficient without manual source code changes, including while applications run in production.
• doubleAI raised $200 million in a Series A round co-led by Lightspeed and Bessemer, with Nvidia and Dell participating.
• According to doubleAI, Maxwell can optimize running programs without developers manually changing their source code.
• The company plans to sell Maxwell to enterprises seeking to reduce AI workload hardware usage.
Read more
According to doubleAI, Maxwell can make software more efficient without manual source code changes, including while applications run in production.
• doubleAI raised $200 million in a Series A round co-led by Lightspeed and Bessemer, with Nvidia and Dell participating.
• According to doubleAI, Maxwell can optimize running programs without developers manually changing their source code.
• The company plans to sell Maxwell to enterprises seeking to reduce AI workload hardware usage.
Read more
Google open-sources TorchTPU for native PyTorch on TPUs
According to Google, TorchTPU lets developers use familiar PyTorch workflows on its TPUs with minimal changes to existing workloads.
• Google has released TorchTPU as open source on GitHub for running PyTorch workloads on its TPUs.
• Google says existing PyTorch workloads need minimal code changes, with support across Hugging Face, serving tools and training frameworks.
• LiteRT now converts PyTorch models for devices, while Google’s TPU agent suite is in early preview on GitHub.
Read more
According to Google, TorchTPU lets developers use familiar PyTorch workflows on its TPUs with minimal changes to existing workloads.
• Google has released TorchTPU as open source on GitHub for running PyTorch workloads on its TPUs.
• Google says existing PyTorch workloads need minimal code changes, with support across Hugging Face, serving tools and training frameworks.
• LiteRT now converts PyTorch models for devices, while Google’s TPU agent suite is in early preview on GitHub.
Read more
Otter.ai raises $320 million from General Catalyst fund
The funding gives Otter.ai capital to expand while it competes in a crowded market for meeting tools.
• Otter.ai said it raised $320 million from General Catalyst’s Customer Value Fund.
• The company plans to use the money for enterprise growth, international expansion and product work.
• Otter.ai says 40 million people have used its tool to date, but it gave no monthly active user or churn figures.
Read more
The funding gives Otter.ai capital to expand while it competes in a crowded market for meeting tools.
• Otter.ai said it raised $320 million from General Catalyst’s Customer Value Fund.
• The company plans to use the money for enterprise growth, international expansion and product work.
• Otter.ai says 40 million people have used its tool to date, but it gave no monthly active user or churn figures.
Read more
Google releases llm-d 1.0 for distributed AI inference
For teams serving their own models, llm-d 1.0 offers ways to reuse cached context and give interactive requests priority when workloads share a fleet.
• Google released llm-d 1.0, an open-source Kubernetes stack for distributed AI inference.
• It adds traffic controls, session-aware routing and coordinated updates for separate model-serving pools.
• Teams can deploy it now on Google Kubernetes Engine with verified Helm charts.
Read more
For teams serving their own models, llm-d 1.0 offers ways to reuse cached context and give interactive requests priority when workloads share a fleet.
• Google released llm-d 1.0, an open-source Kubernetes stack for distributed AI inference.
• It adds traffic controls, session-aware routing and coordinated updates for separate model-serving pools.
• Teams can deploy it now on Google Kubernetes Engine with verified Helm charts.
Read more
Mistral AI releases Voxtral 3 speech models and voice agent SDK
According to Mistral, organizations can build voice agents that complete tasks during calls and choose where audio is processed.
• Mistral AI released Voxtral 3 speech models and an SDK for voice agents that can use tools during live calls.
• Mistral says transcription covers more than 50 languages, while speech generation supports more than 35 languages.
• Voxtral 3 is available now through Mistral’s EU-hosted API or for self-deployment.
Read more
According to Mistral, organizations can build voice agents that complete tasks during calls and choose where audio is processed.
• Mistral AI released Voxtral 3 speech models and an SDK for voice agents that can use tools during live calls.
• Mistral says transcription covers more than 50 languages, while speech generation supports more than 35 languages.
• Voxtral 3 is available now through Mistral’s EU-hosted API or for self-deployment.
Read more
AMD is bringing FSR 4.1 to Xbox consoles and handhelds
FSR 4.1 can improve performance in more than 300 supported games and could have a substantial effect on Xbox games, especially on Series S.
• AMD is bringing FSR 4.1 to Xbox Series S and X, Xbox Ally handhelds, older Radeon GPUs and Ryzen AI devices.
• RX 6000 GPUs and supported Ryzen AI devices start getting FSR 4.1 today through Adrenalin Edition 26.10.1 drivers.
• Xbox developers will get access “very soon,” according to Microsoft’s Jason Ronald; console owners must wait longer.
Read more
FSR 4.1 can improve performance in more than 300 supported games and could have a substantial effect on Xbox games, especially on Series S.
• AMD is bringing FSR 4.1 to Xbox Series S and X, Xbox Ally handhelds, older Radeon GPUs and Ryzen AI devices.
• RX 6000 GPUs and supported Ryzen AI devices start getting FSR 4.1 today through Adrenalin Edition 26.10.1 drivers.
• Xbox developers will get access “very soon,” according to Microsoft’s Jason Ronald; console owners must wait longer.
Read more
Reliable Robotics raises $115 million as valuation tops $1 billion
FAA approval would let Reliable deploy autonomous flight technology across existing aircraft, starting with its own cargo airline.
• Reliable Robotics raised $115 million, bringing its valuation above $1 billion.
• The startup seeks FAA certification for autonomous flight technology designed for existing aircraft.
• Rose expects radar certification by the end of 2027 and initial autopilot certification in 18 to 24 months.
Read more
FAA approval would let Reliable deploy autonomous flight technology across existing aircraft, starting with its own cargo airline.
• Reliable Robotics raised $115 million, bringing its valuation above $1 billion.
• The startup seeks FAA certification for autonomous flight technology designed for existing aircraft.
• Rose expects radar certification by the end of 2027 and initial autopilot certification in 18 to 24 months.
Read more
Perplexity lets subscribers run Computer locally on supported Macs
On supported Macs, local Computer workflows can keep data such as email content off the cloud and give users more control over credit spending.
• Perplexity Pro, Max and Enterprise subscribers can run Computer entirely on a recent Apple Silicon Mac with at least 36GB of RAM.
• The local version includes Computer’s harness and orchestrator and offers Qwen 3.6 or PPLX Pro.
• Users can authorize calls to more than 15 models, while code and tool execution run in an isolated sandbox.
Read more
On supported Macs, local Computer workflows can keep data such as email content off the cloud and give users more control over credit spending.
• Perplexity Pro, Max and Enterprise subscribers can run Computer entirely on a recent Apple Silicon Mac with at least 36GB of RAM.
• The local version includes Computer’s harness and orchestrator and offers Qwen 3.6 or PPLX Pro.
• Users can authorize calls to more than 15 models, while code and tool execution run in an isolated sandbox.
Read more
OpenAI buys Orchestra Bio to expand life sciences tools
The deal adds tools for managing drug discovery operations to OpenAI’s life sciences work, which the company says already serves Moderna and Novo Nordisk.
• OpenAI is buying Orchestra Bio and adding its team and technology to Rosalind and Rosalind Workbench.
• Orchestra makes AI agents that help biotech companies manage drug discovery operations.
• Deal terms were not disclosed; Orchestra had reportedly raised around $20 million as of 2025.
Read more
The deal adds tools for managing drug discovery operations to OpenAI’s life sciences work, which the company says already serves Moderna and Novo Nordisk.
• OpenAI is buying Orchestra Bio and adding its team and technology to Rosalind and Rosalind Workbench.
• Orchestra makes AI agents that help biotech companies manage drug discovery operations.
• Deal terms were not disclosed; Orchestra had reportedly raised around $20 million as of 2025.
Read more
DeepSeek releases DeepInfer inference framework as open source
DeepInfer gives operators access to DeepSeek’s production inference framework, though its cache management and load-balancing code remain unreleased.
• DeepSeek has open-sourced DeepInfer, the inference framework behind its production services.
• It currently supports DeepSeek-V4.1-Flash on Nvidia SM100 Blackwell GPUs and Huawei Ascend 950 chips.
• DeepSeek says one cluster serves 155 trillion tokens daily; it has not released its cache management or load-balancing code.
Read more
DeepInfer gives operators access to DeepSeek’s production inference framework, though its cache management and load-balancing code remain unreleased.
• DeepSeek has open-sourced DeepInfer, the inference framework behind its production services.
• It currently supports DeepSeek-V4.1-Flash on Nvidia SM100 Blackwell GPUs and Huawei Ascend 950 chips.
• DeepSeek says one cluster serves 155 trillion tokens daily; it has not released its cache management or load-balancing code.
Read more