https://zubnet.ai/news/ AI models & releases news by Sarah Chen. en Tue, 07 Apr 2026 09:31:28 +0000 https://zubnet.ai/sarah.png https://zubnet.ai/news/ https://zubnet.ai/news/metas-eupe-vision-encoder-actually-works-your-phone/ https://zubnet.ai/news/metas-eupe-vision-encoder-actually-works-your-phone/ Meta built a sub-100M parameter vision encoder that handles image understanding, segmentation, and VLM tasks without the usual trade-offs. Tue, 07 Apr 2026 04:45:41 +0000 sarah@zubnet.ai (Sarah Chen) Models & Releases https://zubnet.ai/news/metas-open-source-ai-play-faces-reality-check-limited-sources/ https://zubnet.ai/news/metas-open-source-ai-play-faces-reality-check-limited-sources/ Reports claim Meta is planning open-source versions of next-gen models, but sparse details raise questions about the actual scope. Tue, 07 Apr 2026 00:10:31 +0000 sarah@zubnet.ai (Sarah Chen) Models & Releases https://zubnet.ai/news/netflix-void-goes-open-source-physics-aware-video-object-removal/ https://zubnet.ai/news/netflix-void-goes-open-source-physics-aware-video-object-removal/ Netflix's VOID model doesn't just erase objects from videos—it understands physics and fixes what happens when those objects disappear. Sun, 05 Apr 2026 20:25:44 +0000 sarah@zubnet.ai (Sarah Chen) Models & Releases https://zubnet.ai/news/microsofts-three-new-models-still-chasing-openai-independence/ https://zubnet.ai/news/microsofts-three-new-models-still-chasing-openai-independence/ Microsoft quietly released three foundational models through MAI, but are they enough to break free from OpenAI? Fri, 03 Apr 2026 05:25:30 +0000 sarah@zubnet.ai (Sarah Chen) Models & Releases https://zubnet.ai/news/arcees-trinity-large-thinking-apache-20-reasoning-400b-parameters/ https://zubnet.ai/news/arcees-trinity-large-thinking-apache-20-reasoning-400b-parameters/ Finally, an open reasoning model that doesn't come with licensing strings attached. Fri, 03 Apr 2026 00:01:02 +0000 sarah@zubnet.ai (Sarah Chen) Models & Releases https://zubnet.ai/news/microsofts-mai-models-chase-speed-over-innovation/ https://zubnet.ai/news/microsofts-mai-models-chase-speed-over-innovation/ Three new Azure models promise faster voice and image processing, but lack groundbreaking capabilities. Thu, 02 Apr 2026 20:20:31 +0000 sarah@zubnet.ai (Sarah Chen) Models & Releases https://zubnet.ai/news/google-dumps-custom-licensing-gemma-4-local-ai-still-costs-20k/ https://zubnet.ai/news/google-dumps-custom-licensing-gemma-4-local-ai-still-costs-20k/ Apache 2.0 licensing fixes developer complaints, but the hardware reality check remains brutal. Thu, 02 Apr 2026 17:00:55 +0000 sarah@zubnet.ai (Sarah Chen) Models & Releases https://zubnet.ai/news/microsofts-mai-group-drops-three-foundational-models-after-six-months/ https://zubnet.ai/news/microsofts-mai-group-drops-three-foundational-models-after-six-months/ Voice transcription, audio generation, and image models arrive as Microsoft pushes deeper into the foundation model game. Thu, 02 Apr 2026 16:50:30 +0000 sarah@zubnet.ai (Sarah Chen) Models & Releases https://zubnet.ai/news/googles-gemma-4-gets-nvidia-boost-local-ai-agents/ https://zubnet.ai/news/googles-gemma-4-gets-nvidia-boost-local-ai-agents/ New compact models run efficiently on RTX GPUs, targeting the shift from cloud to on-device AI. Thu, 02 Apr 2026 16:20:33 +0000 sarah@zubnet.ai (Sarah Chen) Models & Releases https://zubnet.ai/news/googles-gemma-4-promises-local-reasoning-details-stay-locked-up/ https://zubnet.ai/news/googles-gemma-4-promises-local-reasoning-details-stay-locked-up/ Google claims Gemma 4 matches Gemini 3's architecture for edge reasoning. The devil's in the missing details. Thu, 02 Apr 2026 16:06:43 +0000 sarah@zubnet.ai (Sarah Chen) Models & Releases https://zubnet.ai/news/googles-gemma-4-claims-3-open-model-ranking-whos-counting/ https://zubnet.ai/news/googles-gemma-4-claims-3-open-model-ranking-whos-counting/ Google releases four new Gemma 4 variants with bold performance claims, including a #3 Arena AI ranking that deserves scrutiny. Thu, 02 Apr 2026 16:07:51 +0000 sarah@zubnet.ai (Sarah Chen) Models & Releases https://zubnet.ai/news/ibms-granite-40-3b-vision-targets-enterprise-document-hell/ https://zubnet.ai/news/ibms-granite-40-3b-vision-targets-enterprise-document-hell/ IBM shipped a 3B parameter vision-language model built specifically for chart extraction and table parsing—not general image chat. Thu, 02 Apr 2026 06:05:45 +0000 sarah@zubnet.ai (Sarah Chen) Models & Releases https://zubnet.ai/news/liquid-ais-350m-model-outperforms-1b-models-hybrid-architecture/ https://zubnet.ai/news/liquid-ais-350m-model-outperforms-1b-models-hybrid-architecture/ LFM2.5-350M ditches pure Transformers for a hybrid design that handles 32k context windows while staying under 350M parameters. Wed, 01 Apr 2026 05:05:38 +0000 sarah@zubnet.ai (Sarah Chen) Models & Releases https://zubnet.ai/news/google-rushes-cheap-video-model-sora-shuts-down/ https://zubnet.ai/news/google-rushes-cheap-video-model-sora-shuts-down/ Veo 3.1 Lite costs $0.03/second, half the price of Fast tier, as OpenAI exits standalone video generation. Wed, 01 Apr 2026 03:50:34 +0000 sarah@zubnet.ai (Sarah Chen) Models & Releases https://zubnet.ai/news/googles-veo-31-lite-cuts-video-generation-costs-half-quality/ https://zubnet.ai/news/googles-veo-31-lite-cuts-video-generation-costs-half-quality/ Google released Veo 3.1 Lite at less than 50% the cost of Veo 3.1 Fast, targeting high-volume video apps. The real question: what did they sacrifice? Tue, 31 Mar 2026 16:15:37 +0000 sarah@zubnet.ai (Sarah Chen) Models & Releases https://zubnet.ai/news/alibabas-qwen35-omni-claims-215-sota-wins-native-audio-architecture/ https://zubnet.ai/news/alibabas-qwen35-omni-claims-215-sota-wins-native-audio-architecture/ Qwen3.5-Omni ditches external encoders for a unified 'Thinker-Talker' system trained on 100M hours of audio-visual data. Tue, 31 Mar 2026 05:20:42 +0000 sarah@zubnet.ai (Sarah Chen) Models & Releases https://zubnet.ai/news/microsofts-harrier-models-ditch-bert-decoder-only-embeddings/ https://zubnet.ai/news/microsofts-harrier-models-ditch-bert-decoder-only-embeddings/ New multilingual embedding models abandon traditional bidirectional architecture for LLM-style design with 32k context windows. Mon, 30 Mar 2026 22:25:39 +0000 sarah@zubnet.ai (Sarah Chen) Models & Releases https://zubnet.ai/news/suno-v55-lets-users-train-ai-their-own-voice-music/ https://zubnet.ai/news/suno-v55-lets-users-train-ai-their-own-voice-music/ The music AI company's biggest update yet focuses on personalization over raw quality improvements. Sat, 28 Mar 2026 18:35:37 +0000 sarah@zubnet.ai (Sarah Chen) Models & Releases https://zubnet.ai/news/google-coheres-audio-models-lack-substance-behind-hype/ https://zubnet.ai/news/google-coheres-audio-models-lack-substance-behind-hype/ Two major AI companies dropped audio models this week, but the actual capabilities remain frustratingly vague. Fri, 27 Mar 2026 00:00:53 +0000 sarah@zubnet.ai (Sarah Chen) Models & Releases https://zubnet.ai/news/qwen-35-gets-claude-style-reasoning-easy-deploy-package/ https://zubnet.ai/news/qwen-35-gets-claude-style-reasoning-easy-deploy-package/ New distilled models bring chain-of-thought reasoning to local deployment with GGUF support and 4-bit quantization options. Thu, 26 Mar 2026 23:05:45 +0000 sarah@zubnet.ai (Sarah Chen) Models & Releases