1. Mistral AI Releases Leanstral 1.5 for Lean 4 Code Verification
Mistral AI's new Leanstral 1.5 is a mixture-of-experts model with 119B total parameters and 6.5B active parameters per token. The model features a 256k token context length and supports multimodal inputs. It was trained using a three-stage process consisting of mid-training, supervised fine-tuning, and reinforcement learning with CISPO. In real-world testing across 57 repositories, the model successfully identified 11 genuine bugs, five of which were previously unreported on GitHub. It is available via Hugging Face weights, a free API, or self-hosting with vLLM 0.24.0 or newer.
- • Mistral AI released Leanstral 1.5 under the Apache 2.0 license, available via Hugging Face weights, a free API, or self-hosting with vLLM 0.24.0+.
- • The model is a mixture-of-experts architecture with 119B total parameters and 6.5B active parameters per token.
- • It features a 256k token context length and supports multimodal inputs including text and images.
- • Leanstral 1.5 achieved a 100% success rate on the miniF2F benchmark and solved 587 out of 672 problems on PutnamBench.
- • In real-world testing across 57 repositories, the model identified 11 genuine bugs, including 5 previously unreported on GitHub.
Developers can use this open-weights model or its free API to automate software correctness verification and identify bugs in their codebases.
2. Interfaze Open-Sources Multilingual Audio Diffusion ASR Model
Interfaze has open-sourced diffusion-gemma-asr-small, described as the first open-source multilingual audio diffusion ASR model. It supports English, German, French, Spanish, Hindi, and Mandarin. The model consists of a ~42M-parameter adapter trained on top of a frozen 26B-parameter DiffusionGemma backbone, utilizing a frozen Whisper encoder as a feature extractor. Unlike most diffusion LLMs, it employs uniform, random-token diffusion, meaning transcription cost scales with the number of denoising steps rather than the length of the transcript. The adapter is released under the Apache-2.0 license.
- • Interfaze open-sourced diffusion-gemma-asr-small, a multilingual audio diffusion ASR model supporting English, German, French, Spanish, Hindi, and Mandarin.
- • The model uses a ~42M-parameter adapter trained on top of a frozen 26B-parameter DiffusionGemma backbone.
- • It employs uniform, random-token diffusion and a frozen Whisper encoder as a feature extractor.
- • Transcription cost scales with the number of denoising steps rather than the length of the transcript.
- • The adapter is released under the Apache-2.0 license, while underlying models are subject to their own licenses.
- • Benchmarks on LibriSpeech show it outperforms other diffusion-based ASR models but trails autoregressive Whisper models.
Developers get an open-source, non-autoregressive speech-to-text model where transcription costs scale with denoising steps rather than transcript length.
3. Wafer Runs GLM 5.2 on AMD MI355X at Double the Cost-Efficiency of Blackwell
Wafer achieved an aggregate throughput of 2626 tok/s/node and a single stream performance of 213 tok/s for the GLM5.2 model on AMD MI355X hardware. This inference performance was achieved at over two times lower cost than comparable NVIDIA Blackwell hardware. Wafer utilized the sglang inference framework and quantized the GLM-5.2 model to MXFP4 using AMD Quark. The team implemented two specific fixes in the sglang ROCm image to enable speculative decoding, which resulted in a nearly 3x gain in single-stream throughput, and improved prefill performance by manually tuning the MoE kernel selection for GLM-5.2's fp4 shapes.
- • Wafer achieved an aggregate throughput of 2626 tok/s/node and single-stream performance of 213 tok/s for GLM 5.2 on AMD MI355X.
- • The inference performance was achieved at over two times lower cost than comparable NVIDIA Blackwell hardware.
- • The setup utilized the sglang inference framework and quantized the GLM-5.2 model to MXFP4 using AMD Quark.
- • The team implemented two fixes in the sglang ROCm image to enable speculative decoding, yielding a nearly 3x gain in single-stream throughput.
- • Prefill performance was improved by manually tuning the MoE kernel selection for GLM-5.2's fp4 shapes.
Developers looking to optimize inference costs can leverage AMD MI355X hardware and sglang to achieve high-throughput performance at a fraction of the cost of NVIDIA's latest chips.
4. Laguna XS 2.1 Released for Agentic Coding and Long-Horizon Tasks
Laguna XS 2.1 is optimized specifically for agentic coding and long-horizon tasks. The model achieved a 5.4-point improvement on the SWE-bench Multilingual benchmark, resulting in a score of 63.1%. It is released under the OpenMDW-1.1 license and is available for download on Hugging Face or via an API. To support resource-efficient deployment, the release includes three quantized checkpoints.
- • Laguna XS 2.1 is a 33B parameter Mixture-of-Experts model optimized for agentic coding.
- • The model achieved a 63.1% score on the SWE-bench Multilingual benchmark, representing a 5.4-point improvement.
- • It is released under the OpenMDW-1.1 license and is available on Hugging Face or via an API.
- • Three quantized checkpoints are provided to support resource-efficient deployment.
Developers get a new open-weights coding model option that improves multilingual software engineering performance and offers quantized checkpoints for resource-efficient local deployment.
5. Micro-World Releases Action-Controlled Interactive World Model
Micro-World is an action-controlled interactive world model designed to generate high-quality, open-domain scenes. Built upon the Wan2.1 family of models, the project includes both image-to-world (I2W) and text-to-world (T2W) variants. The T2W model utilizes ControlNet for action injection, while the I2W model uses adaLN for its lightweight parameter footprint. The project has released model weights, full training and inference code, and a curated dataset on GitHub.
- • Micro-World is an action-controlled interactive world model designed to generate open-domain scenes.
- • The model is built on the Wan2.1 family and includes image-to-world (I2W) and text-to-world (T2W) variants.
- • The T2W model uses ControlNet for action injection, while the I2W model uses adaLN for a lightweight parameter footprint.
- • The project has released model weights, full training and inference code, and a curated dataset on GitHub.
Developers can use these open-source weights and training code to build interactive, action-controlled video or scene generation features.
6. Portugal Releases AMALIA 9B Open-Weights LLM
The Portuguese government has announced the release of the AMALIA LLM, a 9B parameter model. The model is available in both SFT and DPO (Direct Preference Optimization) versions on Hugging Face. The AMALIA LLM project is released under an Apache 2.0 license, and a research paper detailing the model is available on arXiv under the identifier 2603.26511.
- • The Portuguese government released the AMALIA LLM, a 9B parameter model.
- • The model is available in both SFT and DPO (Direct Preference Optimization) versions on Hugging Face.
- • It is released under the Apache 2.0 license.
- • A research paper detailing the model is available on arXiv (identifier 2603.26511).
Developers have access to a new Apache 2.0 licensed 9B model optimized with SFT and DPO for local deployment or fine-tuning.
7. Meituan-Longcat Publishes LongCat 2.0 FP8 and INT8 Model Weights
Meituan-longcat has published the model weights for two quantized versions of its LongCat 2.0 model. Both the LongCat-2.0-INT8 and LongCat-2.0-FP8 model weights are now available for download on the Hugging Face platform.
- • Meituan-longcat published the LongCat-2.0-INT8 model weights on Hugging Face.
- • Meituan-longcat also published the LongCat-2.0-FP8 model weights on Hugging Face.
Developers can now download and run quantized versions of the LongCat 2.0 model locally or in their own inference pipelines.
8. pxpipe Proxy Cuts Claude Code Token Costs by Converting Context to Images
pxpipe is a local proxy that reduces Claude Code input token costs by converting bulky context like system prompts, tool documentation, and older history into compact PNG images. The system exploits the fact that image token costs are fixed by pixel dimensions, allowing dense content like code and JSON to be represented more efficiently than as text. It is primarily tuned for the Claude Fable 5 model, which achieves 100/100 on reading benchmarks for imaged content. The tool is lossy, meaning exact verbatim recall for items like IDs, hashes, or secrets is unreliable when stored in imaged content.
- • pxpipe is a local proxy that converts bulky context (system prompts, tool docs, history) into compact PNG images to reduce Claude Code input token costs.
- • The tool exploits fixed image token costs based on pixel dimensions, reducing end-to-end token bills by approximately 59% to 70%.
- • It is primarily tuned for the Claude Fable 5 model, which scored 100/100 on reading benchmarks for imaged content.
- • The tool is lossy, making exact verbatim recall for items like IDs, hashes, or secrets unreliable.
- • In SWE-bench Pro testing, pxpipe-enabled instances achieved a 14/19 resolution rate compared to 15/19 for the control group.
- • Users can deploy the proxy using `npx pxpipe-proxy` and point Claude Code to it via the `ANTHROPIC_BASE_URL` environment variable.
Developers can significantly lower their API bills when using Claude Code by exploiting fixed image token pricing, though exact verbatim recall for secrets is unreliable.
9. Apple Introduces Safari MCP Server for Automated Web Debugging
Apple's Safari Technology Preview 247 introduces the Safari MCP server, a Model Context Protocol server designed for web developers. The server allows MCP-compatible agents to connect to a Safari browser window to debug and troubleshoot web applications. Connected agents can access browser data including the DOM, network requests, screenshots, and console output. The tool operates entirely on the local machine, transmitting captured data directly to the agent rather than to Apple, and does not access personal browser data like AutoFill.
- • Safari Technology Preview 247 introduces the Safari MCP server for web developers.
- • The server allows MCP-compatible agents to connect to a Safari browser window to debug and troubleshoot web applications.
- • Connected agents can access browser data including the DOM, network requests, screenshots, and console output.
- • The tool operates entirely on the local machine, transmitting captured data directly to the agent rather than to Apple.
- • To use it, developers must install Safari Technology Preview and enable remote automation and external agents in browser settings.
Developers can now let AI agents autonomously inspect the DOM, network requests, screenshots, and console logs in Safari to automate web application debugging.
10. Cognition Launches Devin Security Swarm for Whole-Codebase Vulnerability Detection
Cognition has launched Devin Security Swarm, a specialized tool designed to identify security vulnerabilities within complex codebases. The system is built on a novel architecture called Agentic MapReduce, which enables whole-codebase reasoning. The process maps signals across a repository, distributes tasks to agents over bounded shards, reduces findings into a single report, and verifies vulnerabilities in isolated sandboxes. Cognition has published technical documentation detailing the architecture.
- • Cognition introduced Devin Security Swarm to identify security vulnerabilities within complex codebases.
- • The system uses a new "Agentic MapReduce" architecture to perform whole-codebase reasoning.
- • The process maps signals across a repository, distributes tasks to agents over bounded shards, reduces findings into a single report, and verifies vulnerabilities in isolated sandboxes.
- • Cognition has published technical documentation detailing the Agentic MapReduce architecture.
Developers can leverage this agentic system to automatically scan entire repositories, map potential security flaws, and verify them in isolated sandboxes.
11. mcpsnoop Launches as a Live TUI Debugger for Model Context Protocol Traffic
mcpsnoop is a transparent proxy tool designed to debug Model Context Protocol (MCP) traffic between AI clients and servers. The tool operates by sitting in the data path, allowing users to view live JSON-RPC frames in a terminal UI. Key features include hung-call detection, capability inspection, frame search, and the ability to replay captured tool calls. Users can integrate mcpsnoop by wrapping their server command in the client's MCP configuration or by running it as a reverse proxy for streamable-HTTP servers.
- • mcpsnoop is a transparent proxy tool that sits in the data path to debug MCP traffic between AI clients and servers.
- • The tool displays live JSON-RPC frames in a terminal UI and supports frame search, capability inspection, and hung-call detection.
- • It allows developers to replay captured tool calls to test server behavior.
- • Installation is supported via `go install` or Homebrew, and the project is currently in pre-1.0 status.
- • It only executes servers specified in the user's client configuration.
Developers building MCP servers can easily inspect live JSON-RPC frames, detect hung calls, and replay tool calls to streamline debugging.
12. WebBrain Launches as an Open-Source, Local-First AI Browser Agent
WebBrain is a free, open-source AI browser agent for Chrome and Firefox created by Emre Sokullu and licensed under MIT. Operating within the browser's side panel, the agent can run entirely on local models (such as llama.cpp and Ollama) or connect to cloud APIs, with Qwen 3.6 35B recommended for optimal performance. It features two modes: Ask mode for read-only tasks and Act mode for interacting with pages via the Chrome DevTools Protocol. WebBrain is designed with a security-first approach, requiring approval for consequential actions and using the visible UI for mutations instead of direct API calls.
- • WebBrain is a free, open-source AI browser agent for Chrome and Firefox licensed under MIT.
- • The agent operates in the browser side panel and supports local models (llama.cpp, Ollama) or cloud APIs, recommending Qwen 3.6 35B.
- • It features "Ask" mode for read-only tasks and "Act" mode for interacting with pages via the Chrome DevTools Protocol.
- • It uses a security-first approach, requiring approval for consequential actions and using the visible UI for mutations.
- • A managed option, WebBrain Cloud, is available for $5 per month per device profile.
Developers can use or customize this open-source agent to automate browser-based workflows locally using models like Ollama or llama.cpp.
13. SGLang Team Standardizes Agent Workflows into Reusable Playbooks
The SGLang team is standardizing its agent development processes by converting workflows into reusable assets. These assets include SKILL.md files, benchmark contracts, review loops, and production debugging playbooks. The team defines the value of agents as procedural engineering knowledge that is capable of being executed, tested, and reviewed.
- • The SGLang team is converting agent workflows into reusable SKILL.md files and benchmark contracts.
- • The initiative includes review loops and production debugging playbooks.
- • The team defines the value of agents as procedural engineering knowledge that can be executed, tested, and reviewed.
Developers get a structured, executable framework for engineering, testing, and reviewing agentic workflows within the SGLang ecosystem.
14. llama.cpp Adds Experimental Particle Scattering Sampler
The scatter sampler is an experimental feature added to llama.cpp that smooths the next-token probability distribution among top-K candidates. It is designed to reduce generation rigidity without the side effects of increasing temperature, such as pulling in weak tail tokens. The sampler operates after other filters like top-K, top-P, min-P, and XTC have defined the candidate medium, using a Gaussian smoothing kernel based on rank distance to redistribute probability mass among the strongest surviving candidates. It is disabled by default and positioned in the default sampler chain between XTC and temperature.
- • llama.cpp added an experimental "scatter sampler" that smooths next-token probability distributions among top-K candidates.
- • The sampler uses a Gaussian smoothing kernel based on rank distance to redistribute probability mass among the strongest surviving candidates.
- • It operates after other filters (like top-K, top-P, min-P, and XTC) and is positioned between XTC and temperature.
- • Optional features include adaptive strength based on entropy, repeated-token absorption, and collision gating.
- • The sampler is disabled by default and is recommended for creative writing but not for strict coding or math tasks.
Developers can use this sampler to reduce rigidity in creative or roleplay generations while avoiding weak tail tokens, though it is risky for strict coding or math tasks.
15. Tutorial Details Schema-Guided Invoice Extraction Pipeline Using lift-pdf
A new tutorial demonstrates an end-to-end accounts-payable extraction pipeline using lift-pdf to convert synthetic invoice PDFs into structured JSON. The pipeline frames invoice parsing as schema-guided document understanding, targeting fields such as vendor identity, PO numbers, line items, and payment status. The implementation uses GPU-aware model loading with optional 4-bit NF4 quantization and a reusable InferenceManager to optimize performance. Synthetic invoices are generated using ReportLab to ensure mathematical consistency, and the pipeline includes a validation step that calculates field-level accuracy and converts extracted data into a pandas-based accounts-payable ledger.
- • The tutorial demonstrates an end-to-end accounts-payable extraction pipeline using lift-pdf to convert synthetic invoice PDFs into structured JSON.
- • The pipeline targets fields such as vendor identity, PO numbers, line items, and payment status, addressing challenges like separating subtotal from after-tax totals.
- • The implementation uses GPU-aware model loading with optional 4-bit NF4 quantization and a reusable InferenceManager.
- • It includes a validation step that calculates field-level accuracy and converts extracted data into a pandas-based accounts-payable ledger.
- • The tutorial is provided as a reproducible Colab notebook.
Developers can implement this pattern to build highly accurate document intelligence pipelines that handle complex extraction challenges like distinguishing billing addresses.
16. CursorBench 3.1 Released to Evaluate AI Agents on Real-World Tasks
CursorBench 3.1 has been released as a benchmark designed to evaluate the performance of AI agents. The benchmark tests agents using ambiguous, multi-file tasks. All tasks are sourced from real-world Cursor user sessions, providing a realistic evaluation environment for agentic coding tools.
- • CursorBench 3.1 is a benchmark designed to evaluate the performance of AI agents.
- • The benchmark tests agents using ambiguous, multi-file tasks.
- • All tasks are sourced from real-world Cursor user sessions.
Developers can use this benchmark to measure how effectively their coding agents handle complex, multi-file, and ambiguous real-world scenarios.