Ternary-Bonsai-27B and Wan-Dancer-14B trend on Hugging Face, Kimi K3 on AI Gateway, plus agent tools on Product Hunt
TL;DR
Hugging Face models and Google updates lead releases today while Vercel, NVIDIA, and new agent startups add practical tools for inference, robotics, and business workflows.
What shipped
On 16 July multiple model hosts and vendors pushed fresh releases. Hugging Face saw two large models climb the trending list and a paper on compiler feedback for code generation. Google, Vercel, and NVIDIA shipped observability tweaks, new model access, and edge hardware aimed at production use.
Hugging Face trending
Hugging Face hosted three notable additions today. Two large models from prism-ml and Wan-AI rose in visibility while a research paper examined how compilers can steer AI code generation in languages like Rust. The releases give builders immediate access to quantized text and video models plus guidance on safer generation pipelines.
- •Ternary-Bonsai-27B-mlx-2bit prism-ml released a 27B text-generation model quantized to 2 bits with the mlx library. Users can run or fine-tune it directly on Hugging Face for local inference tasks.
- •Wan-Dancer-14B Wan-AI released a 14B image-to-video model built on diffusers. It supports direct download and inference for turning still images into short clips.
- •Generative Compilation paper A new paper shows how Rust compiler feedback can guide AI code generation at each step instead of only after output. The approach reduces errors in strict languages during intermediate token choices.
Vendor launches
Google supplied the largest share of updates with seven product and partnership announcements spanning video tools, search connections, education programs, and policy positions. Vercel added two observability features and Kimi K3 model access while NVIDIA released new Jetson modules for edge robotics.
- •ISR write utilization metric Vercel added a write utilization ratio to ISR Observability so teams can spot routes that write often yet serve few requests and adjust revalidation intervals to cut costs.
- •Runtime cache reasons Vercel now logs why a request missed cache, including time-based or tag-based revalidation, helping debug hit rates for ISR and Partial Prerendering responses.
- •Georgia Public Library AI training Google partnered with the Georgia Public Library Service to offer free Career Certificates and AI courses to residents across the state.
- •Gemini Omni and avatars in Google Vids Google added Gemini Omni and personal avatars to Google Vids so users can generate and star in videos without separate editing tools.
- •Connected apps in AI Mode Google Search will let users securely link third-party services and query them inside AI Mode without leaving the search interface.
- •Gemini Notebook rebrand NotebookLM became Gemini Notebook with tighter Google Workspace integration and a secure cloud environment for note and audio work.
- •Waterloo Futures Lab Google and the University of Waterloo opened a lab where students build production AI tools and prepare for industry roles.
- •Kimi K3 on AI Gateway Moonshot AI's Kimi K3 model with 1M-token context and native vision landed on Vercel AI Gateway for long-horizon engineering and CAD tasks.
- •Bioresilience approach Google DeepMind and Isomorphic Labs published their joint method for building AI models that improve resilience in biological systems.
- •Screwfix AI partnership Google teamed with Screwfix to supply AI tools that help UK tradespeople reduce admin time and expand their businesses.
- •Jetson Thor modules NVIDIA released T3,000 and T2000 modules based on the Thor architecture for running foundation models on mass-market robots and edge devices.
Fal model gallery
Elevenlabs TTS Eleven V3: Elevenlabs released TTS Eleven V3 on Fal for text-to-audio inference. Teams can call it through the HTTP API or playground for immediate voice generation.
Product Hunt picks
Nine AI products appeared on Product Hunt today. Most target agent workflows, private inference, or content automation and give vibe builders ready-made components to test without custom code.
- •ClipMatch ClipMatch turns phone camera rolls into social clips using AI editing and caption tools so creators can publish faster.
- •Graft AI Graft AI converts company operations into a live map that agents can read and act on for internal automation.
- •Zro Zro offers private inference endpoints designed specifically for coding agents that need to keep source code off public clouds.
- •River River supplies AI account executives that run product demos and close B2B deals inside existing sales stacks.
- •Inbix Inbix provides Cloudflare-native email infrastructure aimed at developers who want to send transactional mail from edge functions.
- •Kit For AI Kit For AI adds a persistent memory layer that agents can query across sessions without rebuilding context each time.
- •Verse Verse lets users create and hire autonomous AI employees from a single prompt for repetitive business tasks.
- •The Eureka Database The Eureka Database converts Reddit complaints into validated startup ideas by surfacing repeated pain points.
Replicate new models
faceswap_bare_29: venislathiya814,092 released faceswap_bare_29 on Replicate for raw-to-target face swaps with optional upscale. It accepts direct API calls for batch processing.
Other
Replit, Databricks, LangChain, and AWS each released updates focused on agent reliability and enterprise context handling. The items emphasize measurable gains in code output and search accuracy for production agents.
- •Replit self-driving metrics Replit engineers nearly tripled code output over six months while keeping review times and incident rates flat after adopting AI pair-programming tools.
- •OpenWiki 0.2 LangChain released OpenWiki 0.2 that generates codebase wikis in OKF format with metadata and changelogs agents can retrieve directly.
- •Databricks context engineer certification Databricks launched a Context Engineer Certification plus agent training courses to close the skills gap for building reliable enterprise agents.
- •Unified context layer Databricks described a unified context system that lets AI coworkers pull consistent data across business units instead of working from isolated silos.
- •Amazon Bedrock agent search AWS released Managed Knowledge Bases for Bedrock so agents can run enterprise search over internal documents without custom retrieval code.
- •HeyGen Pictory alternatives HeyGen published a 2026 comparison of eleven video tools that compete with Pictory for AI script-to-video production.
- •HeyGen Submagic alternatives HeyGen published a 2026 comparison of fifteen tools that compete with Submagic for AI caption and subtitle generation.
- •HeyGen Akool alternatives HeyGen published a 2026 comparison of eleven tools that compete with Akool for AI avatar and talking-head video creation.
- •Databricks data foundation warning Databricks noted that most enterprises still lack the clean data layers required before AI agents can operate reliably at scale.
Industry news
Three stories covered macro effects of AI growth. Energy IPOs accelerated, a Codex file-deletion bug was detailed, and Anthropic called for faster state-level AI rules.
- •Energy IPO surge Energy companies tied to AI data-center demand filed IPOs at the fastest pace this century as investors seek exposure to power infrastructure.
- •Codex file deletion bug A reported GPT-5.6 Codex issue deleted user home directories when full access mode ran without sandboxing or auto-review enabled.
What this means for you
For Vibe Builders: You can now test quantized 27B text and 14B video models from Hugging Face without managing servers. Product Hunt launches like Graft AI and Verse give you ready agent maps and prompt-built employees to plug into daily workflows. Watch Kimi K3 on Vercel AI Gateway for long-context tasks that mix code and images.
For Non-techies: Google added Gemini avatars and connected apps to Search and Vids so you can create videos or query linked services without switching tools. Free AI training through Georgia libraries and Screwfix partnerships shows how small teams can adopt these features quickly. New agent tools on Product Hunt handle social clips and admin tasks with one-click sign-up.
For Developers: Vercel added write-utilization metrics and cache-reason logs to help tune ISR and edge caching in production. NVIDIA Jetson Thor modules and Kimi K3 on AI Gateway give concrete options for edge robotics and 1M-token visual reasoning workloads. Replicate and Fal endpoints plus OpenWiki 0.2 let you benchmark new models and keep agent docs machine-readable.
What to watch next
Track next-week releases of any follow-up quantized models from prism-ml or Wan-AI. Watch for production benchmarks on Kimi K3 and Jetson Thor in robotics pilots. Monitor state AI bills after Anthropic's latest comments.
Harsh’s take
The day shows a split between flashy model drops and incremental infrastructure fixes. Most user-facing wins come from Google and small startups while core reliability tooling still trails. Second-order risk is that teams will adopt new models faster than they harden caching and sandbox rules, repeating the Codex deletion pattern at larger scale.
Builders should pick one new endpoint, such as Kimi K3 or Eleven V3, and run a controlled test with existing cache settings this week before scaling.
by Harsh Desai
Sources
Hugging Face trending
- •Ternary-Bonsai-27B-mlx-2bit by prism-ml trends on HuggingFace
- •Wan-Dancer-14B by Wan-AI trends on HuggingFace
- •Generative Compilation: On-the-Fly Compiler Feedback as AI Generates Code
Vendor launches
- •Write utilization now available in ISR Observability
- •Runtime logs now show cache reasons
- •We’re partnering with the Georgia Public Library Service for no-cost career and AI training.
- •Create, edit and star in videos with two Google Vids updates
- •Connect more of your apps to Search
- •NotebookLM is now Gemini Notebook
- •Reimagining higher education with the Waterloo Futures Lab
- •Kimi K3 is now available on AI Gateway
- •Sharpen the Sword, Skip the Downloads: ‘Onimusha: Way of the Sword’ Is Coming to GeForce NOW
- •The DMA should not undercut security & privacy for Europeans
- •Our approach to bioresilience
- •We’re partnering with Screwfix to help the nation's tradespeople nail admin and grow their businesses using AI.
- •NVIDIA Introduces New Jetson Thor Computers to Advance Mainstream Robotics and Edge AI
Fal model gallery
Product Hunt picks
Replicate new models
Other
- •The Self-Driving Company
- •Published11 Best Pictory Alternatives & Competitors Picked For 2026
- •PublishedJuly 15th, 202,615 Best Submagic Alternatives & Competitors Picked For 2026
- •OpenWiki 0.2 brings OKF to codebase documentation
- •Published11 Best Akool Alternatives & Competitors Picked For 2026
- •Your AI is ready. Your data foundation probably isn’t
- •The skills gap behind agentic AI, and how Databricks is closing it with a new context engineer certification and agent trainings
- •Unified context: The missing layer for enterprise AI coworkers
- •Build enterprise search for agents with Amazon Bedrock Managed Knowledge Base
Industry news
More AI news
- Model ReleaseKimi releases K3 open model with 2.8T parameters and 1M context
Kimi launches K3, a multimodal open-weight model with 2.8 trillion parameters and 1M token context. It nears GPT-5.6 Sol and Claude Fable 5 in benchmarks, with full weights due by July 27.
- Model ReleaseGemini adds AI avatars to Google Vids for personalized video creation
Google adds personalized AI avatars to Vids that let users create videos starring digital versions of themselves, alongside Gemini Omni-powered tools for generating and editing videos from prompts and reference images.
- FeaturePerplexity introduces secure sandboxes for safely running AI agents
Perplexity introduces secure sandboxes for agents that provide isolated environments to safely execute code and run complex tasks.