Thinky Releases 975B Open Multimodal Model Inkling Under Apache 2.0Moonshot AI Unveils 2.8 Trillion Parameter Kimi K3, Open Sourcing by July 2026Mira Murati's Thinking Machines Lab Releases First Open-Weight Model InklingReplit engineers tripled code output in 6 months using AI, quality improvedVercel Runtime logs now show cache reasons to debug missesFirefox compiled to WebAssembly runs inside another browserGoogle Vids Adds AI Video Generation and Personal Avatars Powered by GeminiOpenAI launches safer ChatGPT for teens with parental controls and learning toolsGoogle AI Mode lets you securely link more third-party apps to SearchLM Studio launches Bionic, an AI agent for open-source modelsGerman AI Consortium Releases Open 30B Model Soofi S, Tops English & German BenchmarksNew paper proposes self-correcting Markov jump process for joint image understanding and generationGFlowRL: Scaling Distribution-Matching RL to Large Language ModelsAdaptive KV Cache Filtering Corrects Structural-Role Bias in LLM InferenceBlack-Box Audit Tests Whether LLM Chain-of-Thought Relies on Stated PremisesShortOPD Recovers Pruned LLMs' Generation Ability via Short-to-Long DistillationMeta-Learning Preferences for Multilingual LLM Alignment in Low-Resource LanguagesTrain a Gen AI Kick Drum Model on a 6GB VRAM Linux Desktopmermaid-ascii: Convert Mermaid Diagrams to ASCII Art$100 AI Music Video Face-Off: Claude Fable 5 vs GPT-5.6 SolThinky Releases 975B Open Multimodal Model Inkling Under Apache 2.0

Thinky released its first full-size LLM, Inkling, a 975B total parameter, 41B active parameter mixture-of-experts multimodal model under the Apache 2.0 license, trained on 45 trillion tokens of text, images, audio, and video. It also released Inkling-Small (276B total, 12B active). Both models are open-weight.

Chinese AI lab Moonshot AI announced Kimi K3, its most capable model with 2.8 trillion parameters. It’s available now via the website and API, with an open-weight release promised by July 27, 2026. Moonshot calls it the first ‘open 3T-parameter’ model.

Thinking Machines Lab, founded by former OpenAI CTO Mira Murati, released its first open-weight model, Inkling. It’s a 975B total, 41B active parameter mixture-of-experts Transformer under Apache-2.0, trained on 45 trillion tokens of text, images, audio, and video with multimodal capabilities.

Vercel has added a ‘cache reason’ field to Runtime logs, explaining why a request missed the cache, such as time-based or tag-based revalidation. Developers can use this to debug cache misses and improve hit rates. It works for all CDN-cacheable responses, including ISR and Partial Prerendering.

A developer named Puter has compiled Firefox/Gecko to WebAssembly, enabling a full browser to run inside another browser. Leveraging Gecko’s strong single-process support, users can now open a Firefox instance running in WebAssembly inside Chrome and browse the web normally.

Google Vids rolled out two updates: Gemini Omni for AI-powered video creation and editing from text prompts, and personal avatars that let users star in videos using their digital twin. The updates make video production significantly easier.

OpenAI announced enhanced safety features for ChatGPT aimed at teens, including age-appropriate protections, learning tools, parental controls, and expert partnerships. The goal is to provide a safe AI learning environment for teens while giving parents effective oversight.

Google announced that its AI Mode will now allow users to securely link more third-party apps and services, interacting with them directly from search results. This eliminates the need to leave the search interface, improving integration between search and workflows.

LM Studio has launched Bionic, an AI agent designed specifically for open-source models. Users can run open models locally or in the cloud and use Bionic to automate tasks. The product received 131 points and 53 comments on Hacker News.

The German AI consortium released Soofi S, an open 30B parameter model that tops benchmarks in both English and German. The open-weight model is available for download and use, demonstrating strong performance in bilingual settings.

A new paper on arXiv proposes Self-Correcting Coupled Markov Jump Processes, enabling AI systems to perform image understanding and generation simultaneously, breaking the traditional separation of the two. Inspired by how a teacher talks and draws together, the method allows both modalities to correct and reshape each other.

A new paper introduces GFlowRL, applying Generative Flow Networks (GFlowNets) to RL training for LLMs. Instead of maximizing a single reward, GFlowRL matches the reward distribution to encourage diverse reasoning paths, avoiding collapse to a dominant mode, especially useful for large reasoning models.

A new study reveals that existing attention-based KV cache eviction methods (e.g., H2O) suffer from structural-role bias in schema-dense inputs, causing critical tokens to be incorrectly discarded. Researchers propose an adaptive filtering strategy that diagnoses signal energy distribution to dynamically retain important tokens, improving long-context model inference accuracy without additional computational overhead. The technique can be directly applied to optimize reasoning in mainstream LLMs.

A new study introduces Interventional Grounding Audits, a black-box, step-level test that checks whether an LLM’s chain-of-thought (CoT) genuinely depends on its stated premises by substituting predicates during reasoning. Experiments reveal that many seemingly logical CoT chains suffer from ‘premise drift’—the model does not actually use the assumptions it claims to rely on. The tool helps developers evaluate the trustworthiness and consistency of LLM reasoning.

A new study proposes ShortOPD (Short-to-Long On-Policy Distillation) to address the collapse of pruned LLMs on free-form generation tasks. Using short-to-long distillation, it recovers generation quality from compressed checkpoints with only short-sequence training. Experiments show that pruned LLMs restored via ShortOPD achieve performance close to or exceeding the original model across multiple generation benchmarks, while being significantly smaller.

A new paper proposes a meta-learning approach to align multilingual LLMs where low-resource languages lack human preference data. It learns preference meta-knowledge from high-resource languages and transfers it to low-resource ones, significantly improving alignment without requiring extensive labeled data.

A detailed tutorial explains how to train a generative AI kick drum diffusion model on an old Linux desktop with only 6GB VRAM. It covers the full pipeline from dataset preparation and model selection to training optimization, focusing on breaking VRAM limits via reduced precision and parameter tuning. Ideal for indie musicians and developers entering AI audio generation on a budget.

Developer Simon Willison highlights mermaid-ascii, an open-source Go library that converts Mermaid diagrams into plain-text ASCII art. It offers a more complete feature set than an earlier Rust version and has been compiled to WebAssembly via Claude. Ideal for terminals, logs, and code comments where rendered diagrams are not supported.

A comparison article pits Claude Fable 5 against GPT-5.6 Sol in a $100 AI music video challenge, evaluating creativity, consistency, and visual quality. The post scored 92 points and 102 comments on Hacker News.