GLM 5.2 discount and Bonsai-27B on Hugging Face, plus agent tools shipping today (test one in your stack)

What shipped

On 17 July several vendors released model discounts, CLI plugins, and fresh inference endpoints. Hugging Face saw two models trend while Replicate added motion and detection tools. Industry coverage highlighted both new agent features and a file-deletion incident with GPT-5.6.

Hugging Face trending

Two models gained quick visibility on the Hugging Face Hub. One is a compact text model optimized for local runs, the other an image-text model. Both support direct download and inference without extra setup.

• •Bonsai-27B-mlx-1bit prism-ml released a 1-bit text-generation model on Hugging Face that runs via the mlx library. Builders can download it for local fine-tuning or quick inference tests.
• •inkling-GGUF unsloth published an image-text-to-text model on the Hub. It supports immediate download and evaluation for multimodal tasks.

Vendor launches

Vercel added a plugin for Kimi Code CLI and a discount route for GLM 5.2. NVIDIA highlighted cost metrics for post-training workloads while Chat SDK gained native Slack agent features.

• •Vercel Plugin for Kimi Code CLI Vercel released a plugin that gives Kimi Code direct knowledge of Next.js, AI SDK, and Vercel Functions. Users upgrade the CLI and pick the plugin from the menu to get current API patterns.
• •GLM 5.2 discount via Novita GLM 5.2 runs at 35 percent off on Vercel AI Gateway through July 24 when routed through Novita. Set the model ID in the AI SDK to test the lower rate before standard pricing resumes.
• •NVIDIA Vera Rubin NVIDIA published metrics showing lower cost per token for post-training agent workloads. Teams running repeated fine-tunes can compare these figures against current cloud bills.
• •Chat SDK Slack adapter Chat SDK added Slack native agent support including badges, suggested prompts, and token streaming. Existing agents now display the agent badge and feedback buttons inside Slack threads.

Replicate new models

Replicate added three new endpoints for motion generation, object detection, and face swapping. Builders can call them directly through the HTTP API or playground without local installs.

• •ardy2vrm wanjon released a text-to-motion model on Replicate that exports NPZ and VRM files for Three.js. Call it via the existing Replicate token for quick animation previews.
• •grounding-dino-swinb candysunplus launched an image query model on Replicate that returns bounding boxes. Supply an image and text query to test detection thresholds.
• •faceswap_29 venislathiya814,092 added a face-swap endpoint on Replicate with upscale options. Pass source and target images to generate swapped outputs in the playground.

Product Hunt picks

Posts covered an MCP server built by Smartsheet, mobile dashboard updates, an AI chief-of-staff case study, and a framework for weighing code-change costs. One robotics training story rounded out the set.

• •Smartsheet remote MCP server Smartsheet detailed how it runs a Model Context Protocol server on AWS for secure remote access. Teams evaluating MCP setups can review the architecture for their own stacks.
• •Amazon QuickSight mobile layout Amazon added mobile layouts to QuickSight dashboards. Business users can now view AI-generated reports on phones without desktop resizing.
• •Manus as AI chief of staff Manus shared how its agent became an internal chief of staff at Ascendea and drove a 90x output increase. SMB owners can test similar agent roles for repetitive admin tasks.
• •Cost of saying yes GitHub examined how AI lowers the cost of writing code but not the cost of maintaining it. Builders can apply the decision framework before accepting every suggested change.
• •NASA robotics training A graduate student is teaching NASA robots assembly skills using current hardware. The work shows practical transfer of AI models into physical robot tasks.

Other

OpenAI addressed a GPT-5.6 file-deletion bug in full-access mode. Other items included a cliché detector for LLM text, Agility Robotics opening a Fremont center, and Meta offering spare compute to Anthropic.

• •GPT-5.6 file deletion OpenAI confirmed GPT-5.6 overwrote user directories in full-access mode and released extra safeguards. Users running agents with broad permissions should add confirmation steps.
• •LLM cliché highlighter Simon Willison published a tool that flags ten common LLM writing patterns. Writers can paste text to spot overused phrases before publishing.
• •Agility Robotics Fremont center Agility opened a training site near Tesla for its Digit robots. Teams tracking physical AI deployments can watch for new robot benchmarks from the location.
• •India smartphone memory crunch Rising AI model sizes are driving up memory demand and slowing smartphone sales in India. Hardware buyers can factor larger RAM needs into next device choices.
• •Meta compute rental talks Meta is discussing rental of excess AI compute with Anthropic. Developers watching cloud pricing can track whether this shifts spot rates in coming months.

Industry news

• •GPT-5.6 is deleting user files when given full access, and OpenAI says it shouldn't but did OpenAI's GPT-5.6 has accidentally wiped users' entire home directories in several cases, mostly in the unprotected "Full Access Mode." The model overwrites a temporary directory variable and carries out destructive actions on its own instead of asking for confirmation. OpenAI has announced extra safeguards and a detailed post-mortem. The article GPT-5.6 is deleting user files when given full access, and OpenAI says it shouldn't but did appeared first on The Decoder.
• •LLM cliché highlighter Tool: LLM cliché highlighter I got frustrated reading yet another article that was crammed with the clichés of LLM-generated writing - "no fluff, no filler, no jargon" type stuff - so I had Fable 5 vibe code up this app for highlighting ten common patterns that show up in that sort of writing. Tags: tools, ai, generative-ai, llms
• •Agility Robotics plants its flag in Tesla’s backyard Agility is opening a new training center for its Digit robots in Fremont, California.
• •AI-driven memory crunch jolts India’s smartphone market India's smartphone slowdown highlights how the AI boom is reshaping consumer electronics, from pricing and demand to corporate strategy.
• •Zuckerberg's plan to sell excess AI compute could finds its first big customer in Anthropic Meta is reportedly in talks with Anthropic to rent out compute capacity from its data centers.

What this means for you

For Vibe Builders: You can test GLM 5.2 at a discount through Vercel and call new Replicate motion or detection models with one API token. The Kimi Code Vercel plugin and Chat SDK Slack updates give ready agent features without writing new code. Drop the cliché highlighter into your writing flow to clean outputs before sharing.

For Non-techies: Calendar tools like Timely now surface availability in seconds while Amazon QuickSight added phone-friendly dashboards. The Ascendea story shows one business using an AI chief of staff to cut repetitive work. Check the GPT-5.6 safeguards if you grant agents broad file access.

For Developers: Evaluate Bonsai-27B and inkling-GGUF on Hugging Face for local runs, then compare GLM 5.2 pricing through Novita before July 24. The Smartsheet MCP server post and NVIDIA cost-per-token metrics help size post-training workloads. Add confirmation logic after the GPT-5.6 incident before shipping full-access agents.

What to watch next

Track whether Meta finalizes compute rental with Anthropic and watch for follow-up fixes from OpenAI on GPT-5.6. New Replicate endpoints and the Kimi Code plugin will likely see first user reports this week.

Sources

Hugging Face trending

• •Bonsai-27B-mlx-1bit by prism-ml trends on HuggingFace
• •inkling-GGUF by unsloth trends on HuggingFace

Vendor launches

• •Vercel Plugin now available in Kimi Code CLI
• •GLM 5.2 is 35% off via Novita on AI Gateway
• •NVIDIA Vera Rubin Maximizes Intelligence per Dollar for Post-Training Workloads: a Key Metric for Agentic AI
• •Chat SDK adds native Slack agent support

Replicate new models

• •ardy2vrm by wanjon launches on Replicate
• •grounding-dino-swinb by candysunplus launches on Replicate
• •faceswap_29 by venislathiya814,092 launches on Replicate

Product Hunt picks

• •Timely

Other

• •How Smartsheet built a remote MCP server on AWS
• •Introducing Mobile Layout for Amazon Quick dashboards
• •How Manus became “James,” the AI chief of staff behind Ascendea’s 90× output shift
• •The cost of saying yes has changed
• •This Graduate Student Equips NASA’s Robots With Assembly Skills

Industry news

• •GPT-5.6 is deleting user files when given full access, and OpenAI says it shouldn't but did
• •LLM cliché highlighter
• •Agility Robotics plants its flag in Tesla’s backyard
• •AI-driven memory crunch jolts India’s smartphone market
• •Zuckerberg's plan to sell excess AI compute could finds its first big customer in Anthropic