🎙️ Nemotron 3.5 Transcribe: Promptable Multilingual Speech Intelligence
NVIDIA released an open-weight model consolidating 40-language transcription, speaker diarization, translation, and audio QA into a single checkpoint. Built on a 30B architecture with roughly 4B active parameters, it lets developers guide formatting and vocabulary purely via natural-language prompts.
📊 Technical Spec
Input: Audio and Text (Audio) ➡️ output Text
Parameters: 30B (MoE, 4B active)
License: open weights
💰 Availability
Where to find it: open weights on Hugging Face (audio and voice developers)
Cost: Free, open weights to self-host
💡 Bottom Line: A massive win for local speech pipelines, delivering top benchmark accuracy at a fraction of the compute.
#NVIDIA #Nemotron #NVIDIANemotron3_5Transcribe
NVIDIA released an open-weight model consolidating 40-language transcription, speaker diarization, translation, and audio QA into a single checkpoint. Built on a 30B architecture with roughly 4B active parameters, it lets developers guide formatting and vocabulary purely via natural-language prompts.
📊 Technical Spec
Input: Audio and Text (Audio) ➡️ output Text
Parameters: 30B (MoE, 4B active)
License: open weights
💰 Availability
Where to find it: open weights on Hugging Face (audio and voice developers)
Cost: Free, open weights to self-host
💡 Bottom Line: A massive win for local speech pipelines, delivering top benchmark accuracy at a fraction of the compute.
#NVIDIA #Nemotron #NVIDIANemotron3_5Transcribe
♨️ @AInewModels
📊 Hugging Face Open ASR English leaderboard table showing rank 1 among open-weight models at 5.33% average WER
Topping the charts while keeping active parameters tiny, that is how you do it 🎧
#NVIDIA #Nemotron #NVIDIANemotron3_5Transcribe
Topping the charts while keeping active parameters tiny, that is how you do it 🎧
#NVIDIA #Nemotron #NVIDIANemotron3_5Transcribe
♨️ @AInewModels
🧠 TBC Mischief 1: Brain-inspired video generation
The Biological Computing Co. has released TBC Mischief 1, an image-to-video model built on Wan 2.2 that reportedly runs five times faster using architectural lessons learned from living brain cells. Designed for developers and creative teams, it produces short clips on standard chips at an aggressive price of $0.005 per video second.
📊 Technical Spec
Input: Image (Video) ➡️ output Video
License: proprietary
💰 Availability
Where to find it: available now on fal and OpenRouter (creators and video developers)
Cost: Free, open weights to self-host
💡 Bottom Line: Whether bio-computing is real magic or great PR, 5x speed at half a cent per second is hard to ignore.
#TheBiologicalComputingCo #Mischief #TBCMischief1
The Biological Computing Co. has released TBC Mischief 1, an image-to-video model built on Wan 2.2 that reportedly runs five times faster using architectural lessons learned from living brain cells. Designed for developers and creative teams, it produces short clips on standard chips at an aggressive price of $0.005 per video second.
📊 Technical Spec
Input: Image (Video) ➡️ output Video
License: proprietary
💰 Availability
Where to find it: available now on fal and OpenRouter (creators and video developers)
Cost: Free, open weights to self-host
💡 Bottom Line: Whether bio-computing is real magic or great PR, 5x speed at half a cent per second is hard to ignore.
#TheBiologicalComputingCo #Mischief #TBCMischief1
♨️ @AInewModels
🖱️ Mage-VL-Cowork: Compact Hybrid Computer Agent
Hitting an impressive 66.0% on OSWorld-Verified, this compact 4B model punches far above its weight class against frontier systems. It is purpose-built for end-to-end computer use by dynamically switching between GUI clicks and terminal commands without an external router.
📊 Technical Spec
Input: Text and Image (Multimodal) ➡️ output Text
Parameters: 4B
Context window: 32,768 tokens
License: Apache 2.0
💰 Availability
Where to find it: open weights to download (agent developers and researchers)
Cost: Free, open weights to self-host
💡 Bottom Line: Solid proof that running your desktop doesn't require a bloated 70B parameter monster.
#Microsoft #Mage #Mage_VL_Cowork
Hitting an impressive 66.0% on OSWorld-Verified, this compact 4B model punches far above its weight class against frontier systems. It is purpose-built for end-to-end computer use by dynamically switching between GUI clicks and terminal commands without an external router.
📊 Technical Spec
Input: Text and Image (Multimodal) ➡️ output Text
Parameters: 4B
Context window: 32,768 tokens
License: Apache 2.0
💰 Availability
Where to find it: open weights to download (agent developers and researchers)
Cost: Free, open weights to self-host
💡 Bottom Line: Solid proof that running your desktop doesn't require a bloated 70B parameter monster.
#Microsoft #Mage #Mage_VL_Cowork
♨️ @AInewModels
🎨 Palette: generating slide decks in your design language
IBM Research introduced Palette, a 20B open-weight LoRA adapter that builds full presentation decks in code and outperforms frontier systems on PresentBench domains. It splits the workflow into designer and coder roles backed by an automated check-and-repair harness, built for organizations that need strict design compliance on self-hosted infra.
📊 Technical Spec
Input: Text (LLM) ➡️ output Text and Code
Parameters: 20B
License: open weights
💰 Availability
Where to find it: open weights for self-hosting (developers and enterprise teams)
Cost: Free, open weights to self-host
💡 Bottom Line: A clever and practical pipeline for anyone tired of frontier models ignoring brand guidelines and risking confidential data.
#IBM #Palette
IBM Research introduced Palette, a 20B open-weight LoRA adapter that builds full presentation decks in code and outperforms frontier systems on PresentBench domains. It splits the workflow into designer and coder roles backed by an automated check-and-repair harness, built for organizations that need strict design compliance on self-hosted infra.
📊 Technical Spec
Input: Text (LLM) ➡️ output Text and Code
Parameters: 20B
License: open weights
💰 Availability
Where to find it: open weights for self-hosting (developers and enterprise teams)
Cost: Free, open weights to self-host
💡 Bottom Line: A clever and practical pipeline for anyone tired of frontier models ignoring brand guidelines and risking confidential data.
#IBM #Palette
♨️ @AInewModels
Everyone's building agents that click around a desktop, but zero of them can actually drive across your desk. Time to fix that. This all-in-one Arduino robot car coding kit is made for learning and tinkering: you wire it up, write the code, and watch your own program roll around the room.
📊 Technical Spec
Type: Arduino-based smart robot car, complete automation kit
Use: Coding practice, education, and hands-on robotics projects
💰 Availability
Where to find it: AliExpress (students, hobbyists, and parents of curious kids)
Cost: 207 ₪ (was 414 ₪, 50% off)
Buyer feedback: 100% positive ratings, 19 sold last month
💡 Bottom Line: Before the robots take over, you could at least be the one who programmed them. Half price helps the ambition.
🔗 https://s.click.aliexpress.com/e/_c4nRCH2X
#Arduino #Robotics #STEM
📊 Technical Spec
Type: Arduino-based smart robot car, complete automation kit
Use: Coding practice, education, and hands-on robotics projects
💰 Availability
Where to find it: AliExpress (students, hobbyists, and parents of curious kids)
Cost: 207 ₪ (was 414 ₪, 50% off)
Buyer feedback: 100% positive ratings, 19 sold last month
💡 Bottom Line: Before the robots take over, you could at least be the one who programmed them. Half price helps the ambition.
🔗 https://s.click.aliexpress.com/e/_c4nRCH2X
#Arduino #Robotics #STEM
♨️ @AInewModels
📄 Dharma Field Extractor: Document Parsing to JSON
A specialized 2B parameter model that extracts structured JSON fields directly from any document, outperforming general-purpose frontier LLMs in targeted benchmarks. It handles both schema-directed and open-ended extraction at a fraction of the cost.
📊 Technical Spec
Input: Image (Multimodal) ➡️ output Text
Parameters: 2B
💰 Availability
Where to find it: available to try online (developers and enterprises)
💡 Bottom Line: For domain-specific document extraction, a small targeted model beats general-purpose frontier APIs easily.
#DharmaAI #Dharma #DharmaFieldExtractor
A specialized 2B parameter model that extracts structured JSON fields directly from any document, outperforming general-purpose frontier LLMs in targeted benchmarks. It handles both schema-directed and open-ended extraction at a fraction of the cost.
📊 Technical Spec
Input: Image (Multimodal) ➡️ output Text
Parameters: 2B
💰 Availability
Where to find it: available to try online (developers and enterprises)
💡 Bottom Line: For domain-specific document extraction, a small targeted model beats general-purpose frontier APIs easily.
#DharmaAI #Dharma #DharmaFieldExtractor
♨️ @AInewModels
📊 Document field extraction performance vs frontier LLMs
A tiny 2B model outperforming frontier giants on extraction is peak efficiency 👏
#DharmaAI #Dharma #DharmaFieldExtractor
A tiny 2B model outperforming frontier giants on extraction is peak efficiency 👏
#DharmaAI #Dharma #DharmaFieldExtractor
♨️ @AInewModels
🗣️ Phonon TTS: Fast On-Device Speech
Gradium has released Phonon TTS, a lightweight suite of 31M and 40M parameter text-to-speech models designed to run entirely on-device without cloud dependency. It delivers rapid response times with latency as low as 28ms to first audio while supporting multiple languages.
📊 Technical Spec
Input: Text and Audio (Audio) ➡️ output Audio
Parameters: 31M - 40M
License: Apache 2.0 / CC BY-NC
💰 Availability
Where to find it: open weights to download (developers and teams)
Cost: Free, open weights to self-host
💡 Bottom Line: An ultra-compact and lightning-fast local speech engine that brings private voice interfaces to everyday devices.
#Gradium #Phonon #PhononTTS
Gradium has released Phonon TTS, a lightweight suite of 31M and 40M parameter text-to-speech models designed to run entirely on-device without cloud dependency. It delivers rapid response times with latency as low as 28ms to first audio while supporting multiple languages.
📊 Technical Spec
Input: Text and Audio (Audio) ➡️ output Audio
Parameters: 31M - 40M
License: Apache 2.0 / CC BY-NC
💰 Availability
Where to find it: open weights to download (developers and teams)
Cost: Free, open weights to self-host
💡 Bottom Line: An ultra-compact and lightning-fast local speech engine that brings private voice interfaces to everyday devices.
#Gradium #Phonon #PhononTTS
♨️ @AInewModels
📊 Hugging Face Open TTS Leaderboard 95th-percentile time to first audio (TTFA)
28ms to start speaking, which is honestly faster than most humans can formulate a thought ⚡
#Gradium #Phonon #PhononTTS
28ms to start speaking, which is honestly faster than most humans can formulate a thought ⚡
#Gradium #Phonon #PhononTTS
♨️ @AInewModels
🌐 Cosmos 3 Nano: Interactive World Simulation
NVIDIA expands its Cosmos 3 family with four open-weight 16B parameter checkpoints tailored for interactive world simulation and multi-sensor driving scenes. The models enable synchronized multi-camera and LiDAR video generation alongside interactive streaming simulation for autonomous vehicles and robotics.
🆕 What's New (vs Cosmos3-H-Embodied-Nano)
Architecture: 16B parameter checkpoints based on Cosmos 3 Nano
Capabilities: Synchronized multi-view camera and LiDAR generation
Workflows: Streaming generation for robotics and driving policy testing
📊 Technical Spec
Input: Text, Image, Simulation Controls (Video) ➡️ output Video and LiDAR
Parameters: 16B
License: OpenMDW1.1
💰 Availability
Where to find it: open weights on Hugging Face (robotics and AV researchers)
Cost: Free, open weights to self-host
💡 Bottom Line: NVIDIA delivers serious open-weight simulation foundations for autonomous robotics and vehicle developers.
#NVIDIA #Cosmos #Cosmos3NanoWorldModels
NVIDIA expands its Cosmos 3 family with four open-weight 16B parameter checkpoints tailored for interactive world simulation and multi-sensor driving scenes. The models enable synchronized multi-camera and LiDAR video generation alongside interactive streaming simulation for autonomous vehicles and robotics.
🆕 What's New (vs Cosmos3-H-Embodied-Nano)
Architecture: 16B parameter checkpoints based on Cosmos 3 Nano
Capabilities: Synchronized multi-view camera and LiDAR generation
Workflows: Streaming generation for robotics and driving policy testing
📊 Technical Spec
Input: Text, Image, Simulation Controls (Video) ➡️ output Video and LiDAR
Parameters: 16B
License: OpenMDW1.1
💰 Availability
Where to find it: open weights on Hugging Face (robotics and AV researchers)
Cost: Free, open weights to self-host
💡 Bottom Line: NVIDIA delivers serious open-weight simulation foundations for autonomous robotics and vehicle developers.
#NVIDIA #Cosmos #Cosmos3NanoWorldModels
♨️ @AInewModels
⚡ Gemini Omni 1.2 Flash: Longer Clips at Lower Cost
Google introduces native 15-second video generation and editing with improved visual consistency and 30% lower pricing. It offers a fast, budget-friendly pipeline designed for creators and developers producing video at scale.
📊 Technical Spec
Input: Text (Video) ➡️ output Video
License: proprietary
💰 Availability
Where to find it: announced (creators and developers)
💡 Bottom Line: Cheaper and longer native video generation built for high-throughput production pipelines.
#Google #Gemini #GeminiOmni1_2Flash
Google introduces native 15-second video generation and editing with improved visual consistency and 30% lower pricing. It offers a fast, budget-friendly pipeline designed for creators and developers producing video at scale.
📊 Technical Spec
Input: Text (Video) ➡️ output Video
License: proprietary
💰 Availability
Where to find it: announced (creators and developers)
💡 Bottom Line: Cheaper and longer native video generation built for high-throughput production pipelines.
#Google #Gemini #GeminiOmni1_2Flash
♨️ @AInewModels
🧱 Molmo2-4B-tb: Vision Action Robotics
Operating at roughly 6 Hz, this model pairs high-level action planning with precise coordinate pointing for physical robotic tasks. It is specifically tailored for robotics researchers building closed-loop manipulation systems like block stacking.
📊 Technical Spec
Input: Text and Image (Multimodal) ➡️ output Text
Parameters: 4.5B
License: Apache 2.0
💰 Availability
Where to find it: open weights to download on Hugging Face (robotics researchers and developers)
Cost: Free, open weights to self-host
💡 Bottom Line: A practical open-weights bridge connecting lightweight vision-language models to real robotic arms.
#AllenAI #Molmo #Molmo2_4B_tb
Operating at roughly 6 Hz, this model pairs high-level action planning with precise coordinate pointing for physical robotic tasks. It is specifically tailored for robotics researchers building closed-loop manipulation systems like block stacking.
📊 Technical Spec
Input: Text and Image (Multimodal) ➡️ output Text
Parameters: 4.5B
License: Apache 2.0
💰 Availability
Where to find it: open weights to download on Hugging Face (robotics researchers and developers)
Cost: Free, open weights to self-host
💡 Bottom Line: A practical open-weights bridge connecting lightweight vision-language models to real robotic arms.
#AllenAI #Molmo #Molmo2_4B_tb
♨️ @AInewModels
🛡️ GPT-6 Cyber: Specialized exploit research power
The model brings significantly upgraded exploit research and vulnerability analysis capabilities at a much higher API price point. It is tailored specifically for specialized cybersecurity teams needing top-tier security intelligence.
📊 Technical Spec
Input: Text (LLM) ➡️ output Text
License: proprietary
💰 Availability
Where to find it: available now in the API (security researchers and teams)
💡 Bottom Line: Serious exploit research power, but the steep price tag means it is strictly for high-end security teams.
#OpenAI #GPT #GPT_6Cyber
The model brings significantly upgraded exploit research and vulnerability analysis capabilities at a much higher API price point. It is tailored specifically for specialized cybersecurity teams needing top-tier security intelligence.
📊 Technical Spec
Input: Text (LLM) ➡️ output Text
License: proprietary
💰 Availability
Where to find it: available now in the API (security researchers and teams)
💡 Bottom Line: Serious exploit research power, but the steep price tag means it is strictly for high-end security teams.
#OpenAI #GPT #GPT_6Cyber
♨️ @AInewModels