Bun Adds Shared-Memory Threads, llama.cpp Hits b9744, Cline 3.0.29 Ships
Bun opened a PR to add true shared-memory threads to JavaScriptCore, llama.cpp dropped build b9744, and Cline hit v3.0.29 — a solid week for local AI tooling.
Bun opened a PR to add true shared-memory threads to JavaScriptCore, llama.cpp dropped build b9744, and Cline hit v3.0.29 — a solid week for local AI tooling.
Bun just opened a pull request to add shared-memory threads to JavaScriptCore — the engine that powers Bun's JavaScript runtime. If it lands, Bun scripts will be able to share memory across threads the way you can in C or Rust, which matters a lot for anyone building fast local AI pipelines or data-processing tools without Node.js overhead.
llama.cpp shipped build b9744 this week. The project moves fast — a new build roughly every day — but b9744 is worth noting because it keeps pace with the Qwen3 and LTX model families that are trending on Hugging Face right now. If you run local models, pull the latest build before loading any new GGUF files.
Cline, the AI coding agent that runs inside VS Code, released cli-v3.0.29. Cline is one of the more practical tools for non-developers who want an AI to write and edit code files on their behalf — it works with any OpenAI-compatible endpoint, so you can point it at a local llama.cpp server.
Unsloth published Qwen3.6-27B-MTP-GGUF on Hugging Face — a quantized GGUF version of Qwen's 27B multi-token prediction model. Unsloth's GGUF builds are generally well-optimised for consumer hardware. If you have 16–24 GB of VRAM or a beefy Mac, this is worth a test run.
Lightricks dropped LTX-2.3-22b-IC-LoRA-Ingredients, a LoRA add-on for their LTX-2.3 video model focused on ingredient-style image conditioning. It's a niche release — useful if you're building food or product video tools — but it signals that the LTX-2.3 ecosystem is getting specialised fine-tunes quickly.
OmniVoice from k2-fsa appeared as a new Hugging Face Space. It's a voice synthesis and recognition demo. No paper or model card details were available at time of writing.
make-look-scanned is a small open-source CLI (and browser tool via WASM) that makes clean PDFs look like they were photocopied. It's not AI, but it's genuinely useful: if you're building document-processing pipelines and want to test OCR robustness, this gives you realistic scanned-looking test data in seconds.
Ubisoft co-founder Claude Guillemot died in a plane crash on 20 June 2026. This is not an AI story, but Guillemot was a significant figure in the games industry and the news is widely circulating. CBW covers it here out of respect, not because it changes anything for AI builders.
1. Download Unsloth's Qwen3.6-27B-MTP-GGUF and load it in llama.cpp b9744. Run a simple Q&A loop on a topic you know well and compare output quality to whatever model you currently use locally.
2. Install Cline v3.0.29 in VS Code, point it at a local llama.cpp server, and ask it to build a one-page HTML tool that converts a CSV to a table. No coding required on your part — just describe what you want.
3. Use make-look-scanned to generate 10 scanned-looking PDFs from a document you already have, then run them through any OCR tool (Tesseract, Adobe, or a cloud API) to see how well it handles degraded input. Useful if you're building anything that ingests user-uploaded documents.
Confirmed: Bun opened PR #249 on their WebKit fork to add shared-memory threads to JavaScriptCore. The PR is open — it has not merged yet.
Confirmed: llama.cpp build b9744 and Cline cli-v3.0.29 are tagged releases visible on GitHub.
Confirmed: Unsloth's Qwen3.6-27B-MTP-GGUF and Lightricks' LTX-2.3-22b-IC-LoRA-Ingredients are live on Hugging Face.
Not independently verified by CBW: We have not run benchmarks on Qwen3.6-27B-MTP-GGUF or tested OmniVoice. Quality claims are unverified.
Worth noting: OmniVoice (k2-fsa) had no linked paper or detailed model card at time of writing — treat it as an early demo, not a production tool.
Source: Bun WebKit PR #249 — shared-memory threads — https://github.com/oven-sh/WebKit/pull/249
Source: llama.cpp release b9744 — https://github.com/ggml-org/llama.cpp
Source: Cline cli-v3.0.29 release — https://github.com/cline/cline
Source: Unsloth Qwen3.6-27B-MTP-GGUF on Hugging Face — https://huggingface.co/unsloth/Qwen3.6-27B-MTP-GGUF
Source: make-look-scanned on GitHub (HN Show HN) — https://github.com/overflowy/make-look-scanned
Source: Reuters — Claude Guillemot dies in plane crash — https://www.reuters.com/world/ubisofts-co-founder-claude-guillemot-dies-plane-crash-2026-06-20/
Get tomorrow's best AI project in your email. With a guide that works. Free. No spam.