1. KAT-Coder-V2.5 Benchmark Results and AutoBuilder Details Released
Building on the recent release of the KAT-Coder-V2.5-Dev coding model, KwaiKAT has provided additional technical data, including benchmark scores and details on the AutoBuilder system. The model achieved 94.9 on PinchBench, 65.2 on SWE-Bench Pro, and 60.7 on Terminal-Bench 2.1. Furthermore, the AutoBuilder system, which constructs repository environments for training, improved success rates from 16.5% to 57.2% across 12 programming languages.
- • KAT-Coder-V2.5 achieved 94.9 on PinchBench, 65.2 on SWE-Bench Pro, and 60.7 on Terminal-Bench 2.1.
- • The AutoBuilder system improved environment construction success rates from 16.5% to 57.2%.
- • An audit of the RL training process reduced sandbox infrastructure failures from 16% to under 2%.
- • The model was post-trained on Qwen3.6-35B-A3B using 127,000 SFT examples.
These benchmarks and system details provide developers with a clearer understanding of the model's capabilities and the efficacy of the training infrastructure used to build it.
2. GigaChat Audio 10B Released with Temporal Grounding Capabilities
The GigaChat team has released GigaChat Audio 10B, an audio-native Mixture-of-Experts model designed to process and reason over audio directly. By integrating a Conformer speech encoder with a modality adapter, the model can perform complex tasks like audio question answering, classification, and tool-use. Notably, its temporal grounding capabilities allow it to locate specific events in long audio files and generate summaries with precise timestamps, opening up new possibilities for audio-analysis applications.
- • GigaChat Audio 10B is an audio-native LLM built upon the GigaChat 3.1 Lightning text model.
- • The architecture combines a Conformer speech encoder, a modality adapter, and a Mixture-of-Experts decoder.
- • The model supports audio question answering, classification, tool-use, and temporal grounding skills.
- • Temporal grounding capabilities were trained on the TimeGround-1M dataset to enable timestamped event descriptions and localization.
- • A demo and model weights are available on Hugging Face, with research documented in arXiv paper 2607.10387.
Provides developers with an audio-native model capable of precise timestamped event description and localization in long audio files.
3. Sakana AI Expands Fugu Orchestrator with New Cybersecurity Endpoint
Following the initial launch of its Fugu API orchestrator in June 2026, Sakana AI has expanded the platform with a new specialized endpoint, Fugu-Cyber. This addition is designed for cybersecurity tasks, including writing proof-of-concept exploits and mapping MITRE ATT&CK techniques, and is now available for developers subject to manual review and regional restrictions.
- • Fugu-Cyber was released on July 21, 2026, as a new endpoint for the Fugu orchestrator.
- • The model achieved an 86.9% success rate on the CyberGym benchmark.
- • Pricing is $6 per million input tokens and $36 per million output tokens, with higher rates for contexts over 272K.
- • Access is gated by manual review and is currently unavailable in the EU and EEA.
This expansion provides developers with a specialized tool for automated security telemetry and vulnerability analysis within the existing Fugu orchestration ecosystem.
4. Cursor Bridge Runs Claude Code CLI via Cursor Subscription
Developers looking to leverage the power of the Claude Code CLI without paying for a separate Anthropic API key can now use cursor-bridge. This open-source Rust utility acts as a local proxy, routing Claude Code's agentic requests—including file editing and shell execution—directly through an active Cursor subscription. The tool automatically handles authentication by reading tokens from the macOS keychain or environment variables, offering a seamless, cost-effective way to run terminal-based coding agents.
- • cursor-bridge is an open-source Rust binary released under the MIT license.
- • The tool allows Cursor subscribers to run the Claude Code CLI without incurring additional Anthropic API costs.
- • It automates the proxy lifecycle by reading the Cursor auth token from the macOS keychain or Linux environment variables.
- • The tool provides full access to Claude Code's agent capabilities, including file editing, shell commands, and tool use.
- • Prerequisites include having Cursor installed with the agent CLI authenticated and the Claude Code CLI installed.
Allows developers to run the Claude Code CLI agent using their existing Cursor subscription, avoiding separate Anthropic API costs.
5. World Model Optimizer Launches 'wmo serve' for Agent Cost Reduction
The open-source world-model-optimizer project has introduced wmo serve, a tool designed to help developers optimize and downsize the models powering their AI agents. By analyzing agent execution traces, the tool automatically distills repetitive tasks into smaller, specialized models and manages routing between these custom models and frontier APIs. Developers can spin up a local, OpenAI-compatible endpoint that handles token compaction and continuous training, significantly lowering inference costs while maintaining agent performance.
- • The world-model-optimizer project launched 'wmo serve', an open-source tool designed to continually improve agent-specialized models.
- • The tool routes repetitive tasks to distilled smaller models to improve performance and reduce costs.
- • It utilizes agent traces to perform model distillation, model routing, and token compaction.
- • Users can pass agent traces and an OpenRouter key to create a local, OpenAI-compatible endpoint.
- • A hosted version is available via a waitlist, claiming frontier-quality endpoints at a 40% or greater cost reduction.
Enables developers to cut API costs by automatically routing repetitive agent tasks from frontier models to distilled, task-specific local endpoints.
6. Sentient OS Launches Open-Source Local AI Agent for macOS
Sentient OS has launched as a free, open-source local AI agent designed to run entirely on macOS. By analyzing local files, screenshots, and emails overnight using a custom Gemma 4 model, the system builds a secure knowledge base to proactively automate browser and app tasks via its "Sidekick" feature. For complex reasoning, developers can configure the system to route tasks to cloud-based frontier models, with built-in privacy controls that strip personally identifiable information before data leaves the device.
- • Sentient OS is an open-source, local AI tool that analyzes files, screenshots, emails, and messages to proactively assist users.
- • The software requires Apple Silicon (M1 or newer), macOS 15+, and at least 8 GB of RAM.
- • It runs daily on-device inference at 3 AM using a custom Gemma 4 E4B model on a LiteRT-LM fork.
- • A 'Sidekick' feature automates tasks in apps and browsers based on a local knowledge base.
- • The system supports routing complex tasks to frontier models like Qwen 3.7 35B or Kimi K3 via OpenRouter or LM Studio.
- • Privacy is maintained by keeping raw data local and stripping PII from summaries sent to cloud endpoints.
Provides an open-source, privacy-focused local agent framework that automates browser and app tasks using on-device inference.
7. Ruff v0.16.0 Expands Default Rules and Adds Markdown Code Block Formatting
The Rust-based Python linter and formatter Ruff has shipped version 0.16.0, introducing a massive expansion of its default rule set to 413 rules. A key highlight for AI developers is the new ability to format Python code blocks embedded inside Markdown files, making it easier to maintain clean code snippets in prompts and documentation. The update also improves developer workflows by outputting diffs for proposed fixes by default and adding more granular suppression comments.
- • Ruff v0.16.0 increases the number of active rules enabled by default from 59 to 413.
- • The release introduces the ability to format Python code blocks embedded directly within Markdown files.
- • New suppression comments, ruff: ignore and ruff: file-ignore, allow for granular diagnostic suppression.
- • The check and format --check subcommands now output diffs for proposed fixes by default.
- • Twelve rules have been stabilized and moved out of preview, including missing-copyright-notice.
Streamlines Python linting and formatting, especially for managing embedded code blocks in Markdown files common in LLM documentation and prompts.
8. AST-grep Rewrites Tree-sitter Core in Rust for 30% Speedup
The ast-grep project has successfully rewritten the C core of the Tree-sitter parser framework in Rust, delivering a 30% boost in parsing speed and a 22% improvement in end-to-end performance. By utilizing an arena allocator and removing incremental parsing features, the team also slashed peak memory usage on heavy codebases from over 1 GiB down to just 91.2 MiB. Crucially, the new core maintains full ABI compatibility with existing Tree-sitter grammars, making it a drop-in upgrade for tools that parse codebases for LLM context.
- • The ast-grep project rewrote the C core of the Tree-sitter parser framework in Rust.
- • The new Rust-based core achieves a 30% performance increase in parsing and a 22% increase in end-to-end ast-grep performance.
- • Peak memory usage on the TypeScript stress corpus was reduced from over 1 GiB to 91.2 MiB.
- • The rewrite maintains binary interface (ABI) compatibility with existing Tree-sitter grammars and tools.
- • Support for incremental parsing and native loading of WebAssembly-compiled grammars was removed to optimize for file-snapshot analysis.
Accelerates code parsing and file-snapshot analysis for developer tools and LLM context-gathering pipelines.
9. llama.cpp Merges Vision Support for Minimax-M3
The llama.cpp codebase has officially integrated support for the Minimax-M3 model, including both its vision capabilities and Multi-Scale Attention (MSA) architecture. This merge allows developers to run local, resource-efficient multimodal inference on consumer hardware. By bringing Minimax-M3 into the llama.cpp ecosystem, developers can now build offline vision-to-text applications without relying on external APIs.
- • Vision support for the Minimax-M3 model has been officially merged into the llama.cpp codebase.
- • The integration includes support for Minimax M3 with Multi-Scale Attention (MSA).
- • The update enables local, CPU- and GPU-accelerated multimodal inference for Minimax-M3.
Allows developers running local inference via llama.cpp to deploy multimodal vision workflows using the Minimax-M3 model.
10. Benchmark Compares DeepSeek V4 Flash Across Coding Harnesses
A developer's benchmark testing DeepSeek V4 Flash across three popular coding harnesses—Claude Code, OpenCode, and Pi—has revealed that while all three produce identical code diff quality, their operational strategies differ wildly. Claude Code took nearly four times longer than the fastest harness due to its extensive exploration of the codebase, whereas Pi relied on reasoning and OpenCode delegated tasks. For developers running local models like DeepSeek V4 Flash on vLLM, choosing the right harness can dramatically impact execution latency.
- • DeepSeek V4 Flash was tested on three harnesses—Claude Code, OpenCode, and Pi—using a codebase benchmark.
- • The quality of the code diffs produced was identical across all three harnesses.
- • Claude Code took nearly four times longer than the fastest harness to complete the same tasks.
- • The harnesses used different strategies: Pi focused on reasoning, OpenCode on delegating, and Claude Code on extensive codebase exploration.
- • DeepSeek V4 Flash was run on vLLM at approximately 180 tokens per second during the evaluation.
Helps developers choose the most efficient execution harness for running DeepSeek models on codebase tasks without sacrificing code quality.