Releases / Launches

Terry Tao demonstrates coding agents for legacy app modernization

Mathematician Terry Tao published a blog post showing how modern AI coding agents can be used to update and modernize old software applications. The post demonstrates practical techniques for leveraging AI to refactor legacy code, highlighting the potential for AI-assisted software maintenance.

Turboquant v0.3.0 fixes silent FP16 precision bug on Tesla P100 in llama.cpp

A three-line fix in turboquant v0.3.0 corrects a long-standing bug where llama.cpp's CUDA code forced FP16 math on Tesla P100 GPUs, despite the P100 having fast FP16 hardware. The fix restores correct precision and performance for P100 users running llama.cpp.

↑ Updated Sun, Jul 12, 2026, 05:41 AM — Turboquant v0.3.0 released with fix for P100 FP16 bug.

Kreuzberg document extraction tool renamed to Xberg, current version enters LTS

The author of Kreuzberg, a local document extraction tool, announced its renaming to Xberg due to pronunciation and comprehension issues for non-German speakers. The current version will be placed into long-term support (LTS) to simplify the transition, as renaming the repository and overwriting tags is complex.

Zotfile Agents package agent behavior into portable .zot files

Zotfile Agents allow developers to package agent behavior, skills, requirements, and permissions into a portable .zot file. This enables sharing and reuse of agent configurations across different environments.

Jargo: a Go port of Pipecat conversational-AI framework released on GitHub

A developer released Jargo, a WebRTC-native, audio-first conversational-AI framework for Go, as a port of the popular Pipecat framework. The project is available on GitHub and targets developers building real-time voice AI applications in Go.

Toolnexus: a vendor-neutral tool-calling layer for LLMs, byte-identical across 5 languages

Toolnexus is a small, vendor-neutral library that provides a unified tool-calling interface for LLMs, ported byte-identically across JavaScript, Python, Go, Java, and C#. It treats MCP servers, agent skills, custom functions, HTTP endpoints, shell/file tools, and remote A2A agents as the same callable, emitting schemas in OpenAI, Anthropic, and Gemini formats. The library includes a client with built-in parallel and chained tool-calling loops and supports human-in-the-loop suspend/resume.

Developer open-sources Istar Code v5.0.0 deterministic coding agent kernel

A developer released Istar Code v5.0.0, an open-source deterministic coding agent kernel that enforces consistent behavior across models from 1B to 500B parameters. The kernel uses an explicit INSPECT→PLAN→EXECUTE→VERIFY→REFLECT loop with instruction hierarchy, six task classes, and decision table routing to make AI-assisted development more predictable.

HarnessTrim: a deterministic, benchmarked token-economy layer across Claude Code, Codex & OpenCode

HarnessTrim is a new tool that provides a deterministic, benchmarked token-economy layer across coding agent harnesses like Claude Code, Codex, and OpenCode. It uses idempotent reducers to cut token waste from noisy tool output, model verbosity, thinking tokens, and instruction files, without involving an LLM in the reduction path. The tool coordinates existing single-channel solutions (e.g., Caveman, RTK) behind a unified cross-harness policy, and is cache-aware to avoid touching cacheable prompts.

Vellium v1.0.0 released: security hardening, wallpaper-based themes, JSON chat export and a major desktop stability pass

Vellium, a local-first desktop workspace for writing and roleplay with local LLMs, has reached its first stable release v1.0.0. The release focuses on security hardening, desktop stability, and new features like wallpaper-based themes and JSON chat export.

Voodoo Quant claims 95% KLD improvement over Unsloth Dynamic 2.0 on Qwen3.5 models

A developer released two new GGUF quantizations of Qwen3.5 0.8B and 2B using a technique called Voodoo Quant, which optimizes mixed precision by assigning higher precision to more important parts of the model. The author claims Voodoo Quant beats Unsloth Dynamic 2.0 by 95% in Kullback-Leibler divergence (KLD). The quantized models are available on Hugging Face.

Colibrì 744B-Parameter Model Runs on a Laptop

A new 744-billion-parameter model named Colibrì has been released, capable of running on a laptop. The model's name and parameter count suggest a focus on efficiency despite its large size.

Xiaomi quietly uploads MiMo-V2.5-DFlash weights to Hugging Face

Xiaomi has uploaded the official DFlash weights for MiMo-V2.5-DFlash to Hugging Face. The model has 300B+ parameters and runs at 8-10 tokens/second on 2x24GB cards with offloading; DFlash could double that speed. The MTP head was shared but does not work yet.

OpenAI retires group chats in ChatGPT, hints at DM-style Messages tab

OpenAI is removing group chat functionality from ChatGPT, as indicated by new strings in the latest Android build (version 1.2026.188). The strings suggest a shift to a direct-message-style 'Messages' tab, with features like 'Stay on top of messages' and notifications for new messages. This change impacts users who relied on group chats for collaborative conversations.

Workday Launches Developer Tools for Building, Connecting, and Verifying AI Agents for HR, Finance, and IT

Workday announced new developer tools to build, connect, and verify AI agents for HR, finance, and IT. The tools aim to enable developers to create and integrate AI agents within Workday's ecosystem.

Local Context Router cuts vision-token waste by pre-classifying PDF pages

A developer built LocalContextRouter, a tool that runs locally to classify each PDF page as text, OCR-needed, or image before sending to a multimodal model. This avoids the wasteful practice of rendering every page to an image, which can cost 1,300–4,800 tokens per page versus 400–800 tokens for plain text. The tool never calls a model itself, leaving the final API call to the user's app.

vLLM v0.25.0 released with Model Runner V2 as default and PagedAttention removed

vLLM v0.25.0 is now available, featuring 558 commits from 232 contributors. Model Runner V2 becomes the default execution path for all dense models, with new support for EVS, realtime embeddings, prefix caching for Mamba hybrid models, multimodal-prefix bidirectional attention, and dynamic speculative decoding. The legacy PagedAttention implementation has been removed.

New model with 500K-token context and $2/$6 pricing shifts cost calculus

A model offering a 500,000-token context window at $2 per million input tokens and $6 per million output tokens has been released, drawing attention for its cost-effectiveness. The pricing and context length are seen as significant for applications requiring long-context processing, potentially changing the competitive landscape before benchmark comparisons are even made.

VultronRetriever model family announced at Raise Summit Paris, runs offline on iPhone

Vultron announced the VultronRetriever family of models at Raise Summit Paris, demonstrating Q&A and document embedding fully offline on an iPhone. The VultronRetrieverPrime-8B ranks #1 globally on the MTEB Leaderboard, with 16x smaller index storage and 12x higher throughput than previous 9B-class leaders. The VultronRetrieverCore-4.5B ranks second only to Prime, outperforming models twice its size.

temporal-debug-skill project enables AI agents to time-travel debug via git worktree

A new open-source project called temporal-debug-skill gives AI coding agents the ability to debug code at any point in git history. The skill parses fuzzy time references into commit SHAs, creates isolated read-only worktrees, and auto-cleans up, working with any agent like Claude Code or Cursor.

Community curates 51 specialist subagents for Gemini CLI

A GitHub repository collects 51 production-ready subagents for Google's Gemini CLI, spanning 9 domains. Users can drop these into the .gemini/agents/ directory to enable task-specific delegation by Gemini CLI.

agentsweep CLI scans AI coding agent history files for leaked secrets

A new open-source CLI tool called agentsweep scans history files of AI coding agents (Codex, Cursor, Claude Code, Cline, Aider) for plaintext secrets like API keys, DB URLs, and crypto seed phrases. It uses ~191 detection rules from gitleaks plus a dedicated BIP-39 seed phrase detector, addressing a security risk where pasted secrets persist in agent context and can be re-exposed.

FlaxeoUI launches as a frontend for sd.cpp on Windows and Linux

FlaxeoUI, a new graphical interface built around the recent sd.cpp release, was published on GitHub. It exposes most sd.cpp backend capabilities including image generation, editing, video paths, and model/hardware options, with builds for Windows and Linux.

LingBot World open-weights game engine released, runs on consumer GPU

LingBot World, an open-weights game engine with WASD movement and hotkeys for storms, fire, and lightning, was released with a variant targeting a single consumer GPU. The city renders frame by frame as the player moves, but areas are not remembered when revisited. Weights are publicly available.

Salesforce announces Agentforce 360 for enterprise AI agent deployment

Salesforce announced an upgraded version of its Agentforce platform, Agentforce 360, designed to help enterprises build and deploy AI agents. The announcement underscores Salesforce's continued investment in agentic AI for business use cases, as competition in the enterprise AI space intensifies.

Developer creates standalone SearXNG CLI+MCP for open-source coding agents

A developer built a standalone SearXNG CLI and MCP tool that enables open coding agents (like OpenCode, pi coding agent) to perform agentic web search without relying on proprietary APIs or running SearXNG as a standalone Python service. The tool is portable and harness-independent, addressing a key limitation in open-source coding agents.

Article describes implementing Anthropic's MCP in Spring Boot for enterprise AI data access

A technical article explains how to use Anthropic's Model Context Protocol (MCP) with Spring Boot to expose corporate databases to AI agents, replacing custom API wrappers. The approach aims to standardize enterprise AI agent integration with internal data sources.

RDXmin tool adds terse persona and context-saving compressor for Claude

A new tool called RDXmin extends Claude's output terseness with a terse persona system (three levels: lite, full, ultra) and a tool-output compressor that shrinks tool results before they enter context, reducing token usage. The compressor strips ANSI codes, keeps head/tail and error lines, and deduplicates, running behind a safety allowlist to avoid affecting Read/Edit/Write operations.

Google’s SynthID embeds digital watermark during AI generation

Google’s SynthID takes a novel approach to AI content detection by embedding a digital watermark directly into generated content during creation, rather than analyzing it after the fact. This shift from post-hoc detection to proactive identification changes the architecture of AI content provenance.

Experian launches agentic AI capability in Ascend Platform for lending lifecycle

Experian introduced a new core capability within its Ascend Platform to help financial services organizations safely scale agentic AI across the lending lifecycle. The feature brings together Experian, client, and partner AI agents within a trusted operating layer for data, decisioning, governance, and control. A survey cited in the announcement indicates 55% of consumers would allow an AI agent to make an autonomous purchase on their behalf.

Meta launches AI business agent for enterprise automation across WhatsApp, Messenger, Instagram

Meta announced a new AI business agent to help companies automate customer operations across WhatsApp, Messenger, and Instagram. The move targets the enterprise AI market, expanding Meta's commercial offerings beyond consumer applications.

Alibaba launches agentic AI tool for enterprises with Slack, Teams integration plans

Alibaba unveiled a new enterprise AI tool on July 11, 2026, designed as an agentic assistant for businesses. The tool is planned to integrate with Slack and Microsoft Teams, signaling Alibaba's push into enterprise productivity AI.

Databricks Genie Ontology auto-builds corporate context layer on Unity Catalog

Databricks released Genie Ontology, a self-improving context layer that scans queries, pipelines, dashboards, and apps to build a living knowledge graph of business definitions on Unity Catalog. It resolves conflicting definitions automatically, addressing the common failure of AI data assistants that lack corporate context.

Founder builds 9-agent marketing fleet on Hyperagent platform

A founder of Lyrafin AI and LyraSec AI used Hyperagent's Founding 500 program to create a multi-agent marketing system. The platform provided $20,100 in AI credits and access to models like Fable 5, Opus 4.8, Sonnet 5, GPT 5.6, and GLM 5.2. The result was a 'Growth Orchestrator' agent plus eight specialist agents for LinkedIn, X, Reddit, email, SEO, community, analytics, and vision-QA.

Anthropic's Claude Tag brings persistent AI teammate into Slack, not meant to replace Claude Code

Anthropic's Lamis Mukta revealed on the AI Native Dev podcast that Claude Tag, an internal side project turned company-wide tool, extends AI collaboration into Slack as a long-running, proactive teammate. It is not a replacement for Claude Code but rather brings the same context and tools developers use into Slack for persistent collaboration.

Hugging Face releases transformers v5.13.1 patch to enable vLLM compatibility

Hugging Face released transformers v5.13.1, a patch focused on enabling compatibility with the latest vLLM release. The update includes defensive handling of legacy layer types, fixes for custom code with new linear layer names, and a fix for _LazyAutoMapping.register with string keys.

Google Cloud publishes comprehensive guide to agentic AI design patterns

Google Cloud's Architecture Centre released a detailed guide on agent design patterns, offering a clear framework for building reliable AI agents at scale. The guide covers what each pattern is, when to use it, and its costs, providing practical guidance for practitioners.

Ollama releases v0.32.0-rc0 with Qwen3.5 support and agent UI

Ollama released v0.32.0-rc0, adding a Qwen3.5 parser and renderer, a warning for old agent models, and a new agent UI. This update enables users to work with Qwen3.5 models and improves the agent experience.

Claude Code auto mode now available without opt-in on Bedrock, Vertex AI, and Foundry

Anthropic's Claude Code tool has made auto mode available by default on Bedrock, Vertex AI, and Foundry, removing the previous opt-in requirement. The update also fixes terminal freezing during long streaming responses, remote managed settings consent issues, and spurious prompt-injection warnings.

OpenAI releases GPT-5.6 with Sol model, claims to outperform Claude Fable

OpenAI released GPT-5.6 on July 10, 2026, featuring a new Sol model that reportedly surpasses Anthropic's Claude Fable on benchmarks. The release was covered by Fireship on YouTube, noting the timing and performance claims.

Vox Director: open-source agent skill automates Vox-style explainer videos from one topic

A new open-source agent skill called Vox Director automates the creation of Vox-style paper-collage explainer videos from a single topic. It runs on Atlas Cloud API and local ffmpeg, handling script, keyframes, motion, voice-over, music, and captions. The project was released on GitHub on July 10, 2026.

LangChain releases v1.3.13 and langchain-openai v1.3.5 with explicit prompt caching

LangChain released versions 1.3.13 and langchain-openai 1.3.5, adding explicit prompt caching support for OpenAI models. This allows developers to mark reusable prompt prefixes for reduced latency and cost. The update also includes a new 'meta' extra for langchain-meta integration and refreshed model profiles.

SGLang v0.5.15 optimizes GLM-5.2 NVFP4 for Blackwell, achieves 500+ tok/s/user on 8x B300

SGLang released v0.5.15 with tuned GLM-5.2 NVFP4 inference on Blackwell GPUs, reaching 500+ tok/s/user on 8x B300 and 450 on 4x GB300 at batch size 1. The update also enables Spec V2 by default, a zero-overhead speculative decoding method that improves end-to-end throughput by 11% via CUDA-graphable draft-extend and fused metadata ops.

Loop.js framework released for loop engineering with skeptical verify agent

Loop.js is a new open-source framework for loop engineering, allowing developers to define a goal and run iterative rounds until a separate, read-only Verify agent confirms completion. It supports terminal, scheduled, and embedded use, with memory persistence across rounds.

MCP servers and Claude Code CLI updated with new features and fixes

On July 10, 2026, the Model Context Protocol (MCP) servers and Claude Code CLI received updates. The MCP servers release v2026.7.10 updated packages for filesystem, time, fetch, and git. Claude Code CLI v2.1.206 added directory path suggestions for /cd, a /doctor check for trimming CLAUDE.md files, and improved git push handling.

OpenAI launches ChatGPT Work with confusing cloud-local split

OpenAI released ChatGPT Work, a new tier that runs conversations in the cloud on web and mobile, while the desktop app can also access local files and apps. At launch, cloud Work conversations are not visible in the desktop Work interface, and desktop threads remain local. The rollout aims to offer a more integrated productivity experience but has drawn criticism for its confusing cross-platform behavior.

OpenAI launches GPT-5.6 family with Sol, Terra, Luna models

OpenAI announced the general availability of the GPT-5.6 family, including flagship model Sol, balanced model Terra, and cost-efficient Luna. Sol achieves state-of-the-art results across coding, knowledge work, cybersecurity, and science, outperforming previous frontier models with fewer tokens and lower cost per dollar. Pricing ranges from $1/$6 per million tokens for Luna to $5/$30 for Sol. The models have a million-token context window and a February 2026 knowledge cutoff.

OpenLive launches as open-source live voice and vision AI assistant

OpenLive is an open-source, on-device voice and vision AI assistant that provides real-time speech and sight capabilities. It serves as an open alternative to proprietary services like ElevenLabs, Gemini Live, and OpenAI Realtime, allowing users to bring their own model.

LangChain Fireworks partner package v1.4.4 released with cached prompt token reporting

LangChain released version 1.4.4 of the langchain-fireworks partner package on July 9, 2026. The update adds reporting of cached prompt token usage and includes dependency bumps and documentation clarifications.

OpenAI Codex CLI v0.144.0 adds usage-limit credit details, writes approval mode, and interactive MCP auth

OpenAI released Codex CLI v0.144.0 on July 9, 2026, introducing usage-limit reset credits with type/expiration display and selection, a new `writes` app-approval mode for granular read/write permissions, and interactive authentication for MCP tools without experimental opt-in. These features improve developer control over API usage and security.

Meta publishes Muse Spark 1.1 evaluation report with self-conversation attractor states

Meta released the Muse Spark 1.1 Evaluation Report, detailing model behavior including 'Attractor States in Self-Conversation' where two copies of the model produce existential statements. A developer created an LLM plugin for the model after preview access.