1. Anthropic Implements Usage Quotas and Expands Access for Redeployed Fable 5 and Mythos 5
Building on the recent lifting of US export controls, Anthropic has now officially redeployed its Claude Fable 5 and Mythos 5 models. Fable 5 is currently available with a 50% weekly usage limit, transitioning to a credit-based model after July 7. Additionally, Mythos 5 access is being expanded for select US organizations, with Anthropic coordinating with the government to broaden availability to domestic and international partners under the Glasswing program.
- • Claude Fable 5 and Mythos 5 are now officially redeployed.
- • Fable 5 usage is capped at 50% of weekly limits until July 7, when it shifts to a credit-based system.
- • Mythos 5 access is being expanded for select US organizations.
- • Anthropic is working with the US government to extend Mythos 5 access to international partners in the Glasswing program.
Developers can now resume production use of both models, though they must account for new usage quotas and updated access pathways for the Mythos 5 model.
2. Google Begins Testing Successor to Gemini 3.5 Flash on LM Arena
Google is testing an upgraded version of its Gemini Flash model on the LM Arena platform, signaling a potential successor to the recently released Gemini 3.5 Flash. As LM Arena testing historically precedes official public releases, this development suggests Google is preparing a new iteration—rumored to be Gemini 3.6 or 4 Flash—for its high-speed, cost-effective model tier.
- • An upgraded Gemini Flash model is currently undergoing testing on the LM Arena platform.
- • This testing follows the recent release of Gemini 3.5 Flash.
- • LM Arena activity historically serves as a precursor to Google's official model rollouts.
- • The new model is expected to offer incremental performance improvements over the current 3.5 Flash version.
Developers currently using Gemini 3.5 Flash for high-volume API traffic should monitor these tests as they indicate the next performance upgrade in Google's low-latency model roadmap.
3. Kimi K2.7 Code Now Generally Available in GitHub Copilot
Following its initial release as an open-weight model on June 12, 2026, Kimi K2.7 Code is now generally available within GitHub Copilot. This integration allows developers to select the model directly in their IDE, with usage billed through Microsoft Azure. The rollout is active for Copilot Pro, Pro+, and Max plans, with administrative policy enablement required for Business and Enterprise users.
- • Kimi K2.7 Code, previously released as an open-weight model, is now available in the GitHub Copilot model picker.
- • Integration is hosted on Microsoft Azure with usage-based billing.
- • Rollout covers Copilot Pro, Pro+, and Max plans across major IDEs including VS Code, Visual Studio, JetBrains, Xcode, and Eclipse.
- • Business and Enterprise users require manual policy enablement to access the model.
This integration provides a streamlined way for developers to access the Kimi K2.7 Code model directly within their existing IDE workflows, moving beyond manual deployment of the open-source weights.
4. Developer Fine-Tunes Gemma-4-31B for Creative Writing and Copywriting
Developer akwin123 has released a fine-tuned version of the Gemma-4-31B-it model optimized specifically for direct-response copywriting and creative writing. Trained via QLoRA SFT on a curated corpus of marketing briefs and real-world ads, the model scored 1657 Elo on a custom EQ-Bench 3 benchmark—a 290-point jump over the base model. The final weights are merged to full bf16, support a 256K context window, and are available on Hugging Face, with the developer recommending disabling thinking mode for optimal creative output.
- • The Gemma-4-31B-it model was fine-tuned using QLoRA SFT on a curated corpus of marketing briefs and real-world ad examples.
- • The fine-tuned model achieved an Elo score of 1657 compared to the base model's 1367 on a custom EQ-Bench 3 benchmark.
- • The model won 24 out of 30 (80%) head-to-head test cases judged blindly by DeepSeek V4 Flash.
- • The final model weights are merged to full bf16, support a 256K context window, and are available on Hugging Face.
- • The developer recommends disabling thinking mode (enable_thinking = false) for optimal output quality.
Developers building marketing or creative writing applications can download open-weight models specifically optimized to reduce generic AI jargon and improve emotional intelligence.
5. Fijik 2.0 350m Released Under Apache-2.0 License
Developer 'Fijik' has released Fijik 2.0 350m, a tiny open-weight model based on the Granite 4 350M architecture. Released under the Apache-2.0 license, the model was continually pre-trained on 6 billion tokens and post-trained on a custom SFT corpus featuring mixed reasoning efforts. Available on Hugging Face in Safetensors and GGUF formats, the model is highly optimized for constrained environments, though the developer recommends pairing it with web search tools due to its small parameter size.
- • Fijik 2.0 350m is based on the Granite 4 350M architecture and released under the Apache-2.0 license.
- • The model was continually pre-trained on 6 billion tokens with an August 2025 knowledge cutoff.
- • It underwent post-training on a custom SFT corpus featuring mixed reasoning efforts.
- • The model is available on Hugging Face in Safetensors and GGUF formats.
- • The developer recommends using the model with web search tools due to its small parameter size.
It provides a tiny, open-licensed model suitable for highly constrained local environments or web-search-assisted tasks.
6. Manufact Launches MCP Cloud Platform for Model Context Protocol Apps
YC-backed startup Manufact has launched a cloud platform dedicated to the Model Context Protocol (MCP) ecosystem. Positioned as a vertical cloud for MCP (analogous to Vercel's relationship with Next.js), Manufact enables development teams to build, test, monitor, and deploy MCP apps and servers. The platform features GitHub integration, preview deployments for experimental branches, automated testing across ChatGPT and Claude, and support for interactive UIs that display custom branding directly inside AI client interfaces.
- • Manufact is a cloud platform designed to build, test, monitor, and deploy Model Context Protocol (MCP) apps and servers.
- • The platform functions as an MCP vertical cloud, offering GitHub integration, preview deployments, and MCP-specific analytics.
- • Manufact supports automated testing across major clients like ChatGPT and Claude.
- • MCP servers can return interactive UIs, allowing companies to display data and branding directly within AI products.
- • The founders previously operated under the name mcp-use and maintain open-source SDKs for MCP.
It provides a Vercel-like deployment and testing workflow specifically for MCP servers, making it easier to build and distribute interactive tools for ChatGPT and Claude.
7. Alibaba Open-Sources Page Agent for In-Browser GUI Automation
Alibaba has open-sourced Page Agent, a TypeScript-first JavaScript library designed for in-page browser automation. Unlike external tools like Selenium or Playwright, Page Agent runs directly inside the browser session, inheriting the user's active cookies, authentication, and session state. By utilizing a 'DOM dehydration' technique to compress the live DOM into a compact text map, the library can automate interfaces using standard text-based LLMs via any OpenAI-compatible endpoint.
- • Page Agent is an open-source, TypeScript-first JavaScript library released under the MIT license.
- • The library runs inside the browser session, inheriting the user's cookies, authentication, and session state.
- • It uses 'DOM dehydration' to convert the live DOM into a compact text map, enabling compatibility with text-only LLMs.
- • The library is model-agnostic and supports any OpenAI-compatible endpoint.
- • It is limited to single-page interactions and relies on prompt-based safety, requiring server-side validation for sensitive actions.
It allows developers to build AI copilots that inherit the user's active session, cookies, and authentication state directly, bypassing external automation tools.
8. Alibaba's SkillWeaver Cuts Agent Token Consumption by 99%
Researchers at Alibaba have developed SkillWeaver, an open-source framework designed to route subtasks to appropriate tools in complex enterprise AI workflows. By employing a retrieve-and-route approach instead of exposing an entire tool library to the LLM, SkillWeaver reduces context window consumption by over 99%. The framework utilizes a three-stage process to generate a Directed Acyclic Graph (DAG) for parallel execution, and introduces a feedback loop that boosts task decomposition accuracy for 7B models from 51.0% to 67.7%.
- • SkillWeaver uses a three-stage process (Decompose, Retrieve, Compose) to create an execution graph for multi-step tasks.
- • The framework introduces Skill-Aware Decomposition (SAD), a feedback loop that refines task decomposition based on retrieved tools.
- • Activating the SAD feedback loop increased decomposition accuracy for a 7B model from 51.0% to 67.7%.
- • The retrieve-and-route approach reduced context window consumption by over 99% compared to exposing the entire tool library.
- • The framework creates a Directed Acyclic Graph (DAG) to map dependencies and enable parallel execution.
It allows developers to build complex agentic workflows with thousands of tools without exhausting context windows or paying massive token costs.
9. Z.ai Launches ZCode Agentic Development Environment for GLM-5.2
Z.ai (formerly Zhipu AI) has launched ZCode, a free desktop application designed to compete with Cursor and Claude Code. Built as an Agentic Development Environment for the newly open-sourced GLM-5.2 model, ZCode runs on macOS, Windows, and Linux. The underlying GLM-5.2 model is a 744-billion-parameter mixture-of-experts model trained entirely on Huawei silicon and released under the MIT license. ZCode also supports third-party models via bring-your-own-key configurations and offers unique remote control integrations with messaging platforms like WeChat and Telegram.
- • ZCode is a free desktop Agentic Development Environment available on macOS, Windows, and Linux.
- • The environment is designed for GLM-5.2, a 744-billion-parameter mixture-of-experts model with a 1M-token context window.
- • Z.ai released GLM-5.2 open-source weights under the MIT license on Hugging Face on June 16.
- • ZCode supports bring-your-own-key configurations for third-party models.
- • The environment features remote control capabilities via WeChat, Feishu, or Telegram to manage coding tasks from mobile.
It provides developers with a free, cross-platform desktop environment to build with GLM-5.2, featuring remote control capabilities via messaging apps and support for third-party models.
10. Open-Source claude-real-video Prepares Video Content for LLM Analysis
The open-source tool `claude-real-video` has been released under the MIT license to help developers prepare video content for LLM analysis. Rather than using fixed time intervals, the Python-based tool extracts frames based on scene changes and a density floor, removing near-duplicate frames to optimize context window usage. It also transcribes audio using Whisper or existing subtitles, outputting a structured folder with frames, transcripts, and a manifest file compatible with multimodal LLMs.
- • claude-real-video is an open-source Python tool released under the MIT license.
- • The tool extracts frames based on scene changes and a density floor rather than fixed time intervals, removing near-duplicates to save context.
- • It transcribes audio using existing subtitles or the Whisper CLI to generate a structured manifest.
- • It supports local files as well as URLs from YouTube, Instagram, and TikTok.
- • Requirements include Python 3.10+, ffmpeg, and ffprobe.
It allows developers to easily feed video content from local files or URLs (YouTube, TikTok) into LLMs by removing duplicate frames and aligning transcripts.
11. Rust-Based CLI Tool 'ctx' Provides Local Memory for Coding Agents
A new Rust-based CLI tool called `ctx` has been released to provide long-term memory for coding agents. By ingesting transcripts and logs into a local SQLite database, `ctx` uses ranked text matching to let agents search past sessions without requiring complex graph databases or external hosted memory APIs. This local setup allows developers to implement history-researching subagents that brief coding tools on past failures and successful workarounds before starting new tasks.
- • ctx is a Rust CLI tool that provides long-term memory for coding agents using a local SQLite database.
- • The tool uses ranked text matching for search, eliminating the need for graph databases or hosted memory services.
- • Developers can implement an Agent History Research Subagent to prepare briefs on past sessions before starting new tasks.
- • The tool helps agents identify and resolve recurring issues by referencing past failures and successful workarounds.
- • It supports generating cleaned session transcripts that can be shared or attached to pull requests.
It allows developers to give their coding agents access to past session histories and workarounds without needing complex graph databases or hosted memory services.
12. Open-Source ghealth CLI Wraps Google Health API for AI Agents
A new open-source CLI tool called `ghealth` has been released as the official successor to the Fitbit Web API. Serving as a wrapper for the Google Health API v4, the tool is distributed as a single Go binary under the Apache 2.0 license. It exposes 40 verified data types from Fitbit, Pixel Watch, and third-party sources as structured JSON. Designed specifically for terminal and AI agent integration, `ghealth` features deterministic exit codes and includes two Agent Skills files to streamline setup and authentication.
- • ghealth is an open-source CLI tool serving as a wrapper for the Google Health API v4 under the Apache 2.0 license.
- • The tool is the official successor to the Fitbit Web API and uses Google OAuth 2.0 for authentication.
- • Distributed as a single Go binary, it exposes 40 verified data types as structured JSON.
- • The CLI features deterministic exit codes and simplified JSON output designed for terminal and AI agent integration.
- • It includes two Agent Skills files to assist with authentication, setup, and documentation of data types.
It allows developers to easily feed structured health and fitness data into AI agents using deterministic exit codes and pre-built Agent Skills files.
13. FriendliAI Offers High-Speed Inference Layer for Coding Agents
FriendliAI has introduced an inference layer optimized specifically for running AI agents. The platform supports several frontier models, including GLM-5.2, MiniMax-M3, and Kimi-K2.7, and claims the fastest output speed for GLM-5.1 on OpenRouter. Backed by a 99.99% uptime SLA, FriendliAI's infrastructure currently powers developer-centric services such as Claude Code, Cursor, Kilo Code, Hermes Agents, and Ollama.
- • FriendliAI provides a dedicated inference layer designed specifically for running AI agents.
- • The platform claims the fastest output speed for the GLM-5.1 model on OpenRouter.
- • Supported frontier models include GLM-5.2, MiniMax-M3, and Kimi-K2.7.
- • FriendliAI offers a 99.99% uptime service level agreement (SLA).
- • The platform powers backend services for Claude Code, Cursor, Kilo Code, Hermes Agents, and Ollama.
Developers building coding agents can leverage a fast, reliable inference provider with a 99.99% uptime SLA to power tools like Claude Code and Cursor.
14. OpenLumara Adds OpenAI-Compatible API Bridge for Local Model Integration
Building on its initial release as a modular agent framework, OpenLumara now includes an OpenAI-compatible API bridge. Operating on port 8000, this bridge connects front-end UIs like KoboldLite or OpenWebUI to local back-end engines such as llama.cpp or KoboldCPP, replacing cloud-centric API conventions with optimized local performance and token-saving features like thinking header collapsing.
- • OpenLumara now features an OpenAI-compatible API bridge running on port 8000.
- • The bridge connects front-end UIs like KoboldLite and OpenWebUI to local backends like llama.cpp and KoboldCPP.
- • It replaces cloud-centric API conventions to improve performance for local models.
- • Includes specific settings to collapse thinking headers to further reduce token usage.
This update enables broader compatibility for the OpenLumara ecosystem, allowing developers to use standard UI tools with local models while maintaining the framework's focus on token efficiency.
15. Thinking Machines Lab and Bridgewater Fine-Tune Open-Weight Model to Outperform Frontier LLMs
Mira Murati's Thinking Machines Lab and hedge fund Bridgewater Associates published joint results evaluating AI for filtering investment news. While frontier models like GPT, Claude, and Gemini topped out in the mid-70s even with expert prompting, Bridgewater fine-tuned an open-weight model using the Tinker API. The resulting model achieved 84.7% accuracy—surpassing the 80% operational trust threshold—while slashing costs by 13.8x compared to the best frontier model.
- • Frontier models (GPT, Claude, Gemini) achieved only ~50% average accuracy across six investment news filtering tests.
- • Expert-written prompts improved frontier model accuracy to the mid-70s, still below Bridgewater's 80% operational trust threshold.
- • Bridgewater fine-tuned an open-weight model using Thinking Machines Lab's Tinker API based on real expert judgment calls.
- • The fine-tuned model achieved 84.7% accuracy, reducing mistakes by 29.8% compared to the best performing frontier model tested.
- • The fine-tuned model delivered a 13.8x lower cost per task compared to the top-performing frontier model tested.
It demonstrates a concrete case where fine-tuning a smaller open-weight model on expert data yields higher accuracy and significantly lower costs than relying on frontier APIs.
16. Hugging Face Developer Automates Role Using Claude Agents SDK and Modal
At the aiDotEngineer event, a Hugging Face developer demonstrated a workflow designed to automate aspects of their role. The implementation combines the Claude Agents SDK with GLM-5.2 accessed via inference providers. For production monitoring and deployment, the setup integrates Langfuse for tracing and runs on Modal's serverless platform, offering a concrete stack reference for developers building autonomous agents.
- • The automation workflow was presented at the aiDotEngineer event by a Hugging Face developer.
- • The system utilizes the Claude Agents SDK and GLM-5.2 via inference providers.
- • Langfuse is used for LLM tracing and monitoring.
- • The entire agentic workflow is deployed on Modal's serverless infrastructure.
It provides a real-world blueprint of a production-grade agentic workflow combining SDKs, tracing, and serverless deployment.
17. RAG-Anything Tutorial Details Multimodal Retrieval in Google Colab
A new tutorial outlines how to build a RAG-Anything workflow for multimodal retrieval involving text, tables, equations, and images. Implemented in Google Colab using OpenAI APIs for chat, vision, and embeddings, the guide details how to convert complex documents into a structured JSON content list. The RAG-Anything system supports naive, local, global, and hybrid retrieval modes, enabling developers to implement cross-modal reasoning over mixed-media datasets.
- • The tutorial demonstrates building a RAG-Anything workflow in Google Colab using OpenAI APIs for chat, vision, and embeddings.
- • The workflow converts synthetic multimodal reports (including tables, charts, and PDFs) into a structured JSON content list.
- • RAG-Anything supports multiple retrieval modes, including naive, local, global, and hybrid.
- • The system is tested on its ability to perform cross-modal reasoning and answer queries based on structured values and multimodal evidence.
It provides a step-by-step guide for developers to implement cross-modal reasoning and retrieval pipelines using standard OpenAI APIs.
18. vLLM Updates Resolve OOM Issues and Double Context Windows
Developers of the open-source vLLM engine have released three major updates aimed at resolving Out of Memory (OOM) issues. The updates address memory bottlenecks caused by preallocations and tuning parameters. In real-world testing, the optimizations allowed a developer to double their usable context window from 120k to 240k tokens when running Qwen2-7B on an RTX 5090 GPU, significantly improving local inference capabilities.
- • vLLM developers released three major updates addressing Out of Memory (OOM) issues caused by preallocations and tuning.
- • A user reported doubling their usable context window size from 120k to 240k tokens using Qwen2-7B on an RTX 5090.
- • The updates optimize memory preallocation and tuning parameters for local inference.
Developers running local models can now double their usable context windows (e.g., up to 240k tokens) on consumer GPUs like the RTX 5090.
19. GLM 5.2 Gains Speculative Decoding Support via DSpark
The GLM 5.2 DSpark preview has been released, marking the first application of the open-sourced DSpark framework to a non-DeepSeek model. By utilizing the framework's speculative decoding capabilities, the preview accelerates FP8 decoding for GLM-5.2 by approximately 1.5x on 4x B300 hardware via vLLM nightly. This release demonstrates the framework's utility for optimizing inference latency across diverse open-weight architectures.
- • The GLM 5.2 DSpark preview is the first speculator developed for a non-DeepSeek model using the open-sourced DSpark framework.
- • It achieves 1.5x faster decoding for GLM-5.2-FP8 on 4x B300 hardware.
- • The implementation runs on vLLM nightly.
- • Stronger checkpoints are expected in future updates.
Developers can now apply the DSpark speculative decoding framework to GLM-5.2 to significantly reduce inference latency, validating the framework's broader compatibility beyond the original DeepSeek-V4 models.
20. llama.cpp Pull Request Doubles Prompt Processing Speeds on Intel ARC
A new pull request for the `llama.cpp` repository introduces significant prompt processing speedups for Intel ARC GPUs. Developed with assistance from Claude, the optimization was tested on an Intel B580 GPU running a Qwen3.6 35B model with a 262k context size. The update cut processing time for a 116k context conversation from 510 seconds to 262 seconds, effectively doubling throughput from 245 t/s to 462 t/s. The current optimization is limited to F16 KV cache.
- • A new llama.cpp pull request provides performance optimizations for Intel ARC GPUs.
- • Testing on an Intel B580 GPU with a 116k context conversation cut processing time from 510 seconds to 262 seconds.
- • Prompt processing speed increased from 245 t/s to 462 t/s.
- • The benchmarks were conducted using the Qwen3.6 35B A3B Q5_K_XL model with a 262,144 context size.
- • The current optimization is limited to F16 KV cache.
Developers running local models on Intel ARC hardware can significantly reduce latency for long-context conversations.
21. Developer Benchmarks 6x NVIDIA P40 Home Lab for MiniMax-M2.7
A developer has documented a home lab configuration designed to run large models locally using six NVIDIA P40 GPUs, providing 144GB of total VRAM. Built on an Asus X99-E-WS motherboard with a modded BIOS and an Intel Xeon processor, the system was benchmarked using llama.cpp to find the optimal configuration for running MiniMax-M2.7-GGUF. The developer's verified setup utilizes F16 KV cache, Flash Attention, a batch size of 2048, and layer-based tensor splitting.
- • The home lab system features six NVIDIA P40 GPUs providing a total of 144GB of VRAM.
- • The hardware includes an Asus X99-E-WS motherboard with a modded BIOS, an Intel Xeon E5-2680 v4 CPU, and 128GB of DDR4 RAM.
- • The optimal llama.cpp configuration uses F16 KV cache, Flash Attention, a batch size of 2048, a ubatch size of 256, and layer-based tensor splitting.
- • The setup was benchmarked running the MiniMax-M2.7-GGUF model via llama-server.
It provides a concrete hardware and software configuration blueprint for developers looking to run large models locally on older, cost-effective enterprise GPUs.
22. Local Benchmarks Compare Qwen3.6, Gemma4, and Ornith1.0 on RTX 3090
An independent developer has published local benchmark results comparing Qwen3.6 27b, Gemma4 26B, and Ornith1.0 35B on an RTX 3090 GPU using the inspect-ai framework. The tests revealed that Qwen3.6 27b led in general knowledge and reasoning, while Ornith1.0 35B excelled in grounding and recall tasks up to 100k tokens. The benchmarks also highlighted practical integration challenges, such as infinite looping issues when running Gemma4 26B via LM Studio.
- • The benchmarks evaluated Qwen3.6 27b, Gemma4 26B, and Ornith1.0 35B using the inspect-ai framework on an RTX 3090.
- • Qwen3.6 27b achieved the highest or joint-highest scores in 4 out of 6 general knowledge and reasoning benchmarks.
- • Ornith1.0 35B outperformed other models in grounding and recall tasks, including a 100k-token Needle in a Haystack test.
- • The evaluator reported technical issues, including infinite looping with Gemma4 26B and long processing times for coding benchmarks.
It provides developers with objective, local hardware performance data to help choose between popular open-weight models for reasoning and recall tasks.