1. Moondream 3.1 Released with 9B Parameter Mixture-of-Experts Architecture
Moondream 3.1 has been released as a vision language model featuring a mixture-of-experts architecture. The model contains 9 billion total parameters with 2 billion active parameters. It is designed for visual reasoning and detection tasks, supporting native skills including query, detect, point, and caption. All skills in Moondream 3.1 return structured output.
- • Moondream 3.1 features a mixture-of-experts (MoE) architecture with 9 billion total parameters and 2 billion active parameters.
- • The model is designed for visual reasoning and detection tasks.
- • It supports native skills including query, detect, point, and caption.
- • All native skills in Moondream 3.1 return structured output.
It provides a lightweight, highly efficient open-weights vision model for local visual reasoning, detection, and structured output tasks.
2. Report Finds Severe Security Vulnerabilities Across Popular MCP Servers
Canopii's 'State of MCP Security 2026' report analyzed 11,524 published MCP servers. The report found that 830 servers, representing one in every fourteen, received a security grade of D or F. Researchers identified 232 servers containing dangerous code sinks, such as eval, shell injection, or unsafe deserialization. The study also found 184 server versions that changed their tool definitions after initial publication. Endpoints claiming to require authentication served tools to anonymous callers in 741 instances. Notably, servers with 1,000 or more GitHub stars are over five times more likely to be high-risk than servers with fewer than 10 stars, with six of the 15 most-starred servers receiving a D or F grade.
- • Canopii's report analyzed 11,524 published MCP servers and graded 830 of them (one in fourteen) with a D or F.
- • Researchers identified 232 servers containing dangerous code sinks, such as eval, shell injection, or unsafe deserialization.
- • The study found 184 server versions that changed their tool definitions after initial publication.
- • Endpoints claiming to require authentication served tools to anonymous callers in 741 instances.
- • Servers with 1,000 or more GitHub stars are over five times more likely to be high-risk than servers with fewer than 10 stars.
- • Six of the 15 most-starred servers in the registry received a security grade of D or F.
Developers integrating third-party MCP servers must carefully audit them for vulnerabilities like shell injection and unauthorized tool exposure to protect their applications.
3. llama.cpp Release Fixes Context Checkpoint Bug in Agentic Workflows
The llama.cpp release b9978 addresses a checkpoint bug that specifically impacted agentic workloads. Prior to the fix, every agent turn created a new checkpoint, which bypassed minimum-step spacing and collapsed the coverage window. The previous bug caused context rewinds during tool-calling loops to erase all checkpoints, resulting in a full reprocess. The update introduces the eviction of closely-spaced checkpoints from prior tasks to maintain a wider coverage window, improving the speed of long agent sessions.
- • The llama.cpp release b9978 addresses a checkpoint bug that specifically impacted agentic workloads.
- • Prior to the fix, every agent turn created a new checkpoint, bypassing minimum-step spacing and collapsing the coverage window.
- • The bug caused context rewinds during tool-calling loops to erase all checkpoints, resulting in a full reprocess.
- • The update introduces the eviction of closely-spaced checkpoints from prior tasks to maintain a wider coverage window.
This fix significantly improves the speed and efficiency of long local agent sessions during complex tool-calling loops.
4. Zer0Fit MCP Server Brings Google's TabFM and TimesFM Models to Local Workflows
A graduate student has released Zer0Fit, an MCP wrapper that allows Google's TabFM and TimesFM transformer models to run in a single Docker container. The tool enables integration of these tabular and time-series models with local LLMs like Open WebUI, Claude Code, and Codex. Zer0Fit requires at least 16GB of VRAM and is limited to CUDA-compatible hardware due to its PyTorch foundation. The implementation features dynamic model loading and unloading with a 5-minute time-to-live (TTL) to manage VRAM usage. Initial testing on classic datasets yielded 94.7% accuracy for the Iris classifier and an R2 score of 0.87 for the California Housing regression model.
- • Zer0Fit runs Google's TabFM and TimesFM transformer models locally in a single Docker container as an MCP server.
- • The tool enables integration of tabular and time-series models with local LLMs like Open WebUI, Claude Code, and Codex.
- • It requires at least 16GB of VRAM and is limited to CUDA-compatible hardware due to its PyTorch foundation.
- • The implementation features dynamic model loading and unloading with a 5-minute time-to-live (TTL) to manage VRAM usage.
- • Initial testing yielded 94.7% accuracy for the Iris classifier and an R2 score of 0.87 for the California Housing regression model.
- • The project currently supports CSV files, with plans to add support for XLS, XLSX, JSON, and JSONL formats.
It allows developers to integrate powerful tabular and time-series forecasting, classification, and regression models directly into local LLM workflows and agent environments.
5. Mindwalk Visualizes Claude Code and Codex Agent Sessions in 3D
Mindwalk is a new visualization tool that replays coding-agent sessions on a 3D map of a codebase. The tool processes session logs from Claude Code and Codex locally without transmitting data off-machine. It represents the repository as a 3D map where file activity is visualized through glowing effects based on search, read, and edit actions. Mindwalk is built with a Go backend and a React/Three.js frontend, utilizing normalized trace and citymap JSON files. It is licensed under the MIT license and is credited to Ricko Yu.
- • Mindwalk processes session logs from Claude Code and Codex locally to generate a 3D map of codebase activity.
- • File activity is visualized through glowing effects based on search, read, and edit actions.
- • Features include tree and terrain views, touch state tracking, a playback deck with a histogram, and timeline marks for user turns and subagent launches.
- • The tool is built with a Go backend and a React/Three.js frontend, utilizing normalized trace and citymap JSON files.
- • Mindwalk is licensed under the MIT license and is available for installation via a shell script or building from source.
It helps developers debug and understand complex multi-step agent behaviors through an interactive 3D visualization of codebase activity.
6. Modelr App Brings Local Image-to-3D Generation to Apple Silicon
Developer Zimeng Xiong released Modelr, a standalone image-to-3D desktop application for Apple Silicon based on a Swift/MLX port of Hunyuan3D-Paint and Hunyuan3D-Shape. The application utilizes the MLX framework to run models on Apple Silicon without the overhead of PyTorch or CPU processing. Benchmarks on an M4 Max chip in FP16 show that hy3d shape (small) runs in 20.9 seconds with 5.6 GB of peak memory, while hy3d paint (pbr) runs in 344 seconds with 39 GB of peak memory. Modelr features include background removal using SwiftVision and real-time diffusion streaming for 3D model generation.
- • Developer Zimeng Xiong released Modelr, a standalone image-to-3D desktop application for Apple Silicon.
- • The application uses the MLX framework to run Hunyuan3D-Paint and Hunyuan3D-Shape without PyTorch or CPU processing overhead.
- • Benchmarks on an M4 Max chip in FP16 show hy3d shape (small) runs in 20.9 seconds with 5.6 GB of peak memory.
- • The hy3d paint (pbr) model runs in 344 seconds with 39 GB of peak memory on the same hardware.
- • Features include background removal using SwiftVision and real-time diffusion streaming for 3D model generation.
- • The source code and weights are available on GitHub under the Hunyuan3D-Swift and Modelr repositories.
It gives developers a fast, lightweight, and local way to generate 3D models from images on Mac hardware without relying on cloud APIs.
7. Native GGUF Jacobian-Lens Visualizer and Steerer Released for llama.cpp
A developer has released an interactive Jacobian-lens visualizer and live steerer designed for GGUF models running on llama.cpp. Inspired by Anthropic's research, the project features a native GGUF server that supports model observation, J-space swapping, abliteration, and steering. The tool is compatible with both dense and Mixture-of-Experts (MoE) GGUF models and can observe running llama-server instances. Memory requirements for the lens scale at approximately 1/8 of the model size, requiring about 20 GB of additional RAM for a 160 GB model.
- • The jlens-gguf tool features a native GGUF server that supports model observation, J-space swapping, abliteration, and steering.
- • The tool is compatible with both dense and Mixture-of-Experts (MoE) GGUF models and can observe running llama-server instances.
- • Memory requirements for the lens scale at approximately 1/8 of the model size, requiring about 20 GB of additional RAM for a 160 GB model.
- • The project code is publicly available on GitHub at github.com/igorbarshteyn/jlens-gguf.
This tool gives developers a concrete way to inspect and steer the internal activations of local GGUF models during inference.
8. mlx-lm Adds Native Support for Nemotron Puzzle 75B on Apple Silicon
Native support for nemotron_h_puzzle was added to the mlx-lm library via pull request #1535. A performance comparison between 4-bit and 5-bit expert quantization was conducted on a 64GB M2 Max, with both configurations using 6-bit dense layers and a BF16 output head. The 4-bit configuration outperformed the 5-bit version, achieving 14.27 tokens per second compared to 10.53 tokens per second. Implementing the model required adding blockwise configuration support, tensor remapping, and matching NVIDIA's FP32 norm and router behavior. A discrepancy in first-layer SSM outputs was resolved by adjusting the promotion order of softplus calculations to match NVIDIA's reference, improving cosine similarity to 0.999998.
- • Native support for nemotron_h_puzzle was added to the mlx-lm library via pull request #1535.
- • A 4-bit expert quantization configuration achieved 14.27 tokens per second on a 64GB M2 Max, outperforming a 5-bit configuration (10.53 t/s).
- • The 5-bit configuration suffered from memory pressure, leading to lower performance in local task checks and long-context retrieval.
- • Implementing the model required adding blockwise configuration support, tensor remapping, and matching NVIDIA's FP32 norm and router behavior.
- • A discrepancy in first-layer SSM outputs was resolved by adjusting the promotion order of softplus calculations, improving cosine similarity to 0.999998.
- • The use of a BF16 output head is mandatory to prevent the 131k-vocabulary lm_head from generating repetitive output.
It allows developers to run a powerful 75B parameter model locally on Apple Silicon with optimized memory usage and performance.
9. DeepSeek Further Widens Cost Gap with 75% V4-Pro Price Cut
Building on the existing cost disparity between DeepSeek and U.S. frontier labs, DeepSeek has now reduced its V4-Pro model pricing by 75%. While this further lowers the barrier for developers, the rapid growth of token consumption in multi-step agentic workflows continues to outpace these falling inference costs. As previously noted, agentic systems can reach input-to-billed token ratios of 1:700, meaning that even with lower unit costs, developers must still adopt techniques like prompt caching and cost-aware routing to maintain sustainable margins.
- • DeepSeek has reduced V4-Pro pricing by 75%, further increasing the cost advantage over U.S. frontier models.
- • Despite lower unit costs, multi-step agent workflows continue to consume tokens at rates that can reach 1:700 ratios.
- • Token amplification remains a primary driver of negative gross margins for vendors using traditional seat-based pricing.
- • Developers are increasingly forced to implement cost-management strategies like prompt caching and speculative decoding to offset high agentic token usage.
While DeepSeek's price cut makes it an even more attractive alternative to expensive U.S. frontier models, the inherent token amplification of agentic workflows means that unit cost reductions alone may not solve the underlying margin challenges for enterprise vendors.
10. Anthropic Extends Fable 5 Promotional Access and Rate Limits Through July 19
Anthropic has extended the promotional period for Claude Fable 5 access and the 50% higher weekly rate limits for Claude Code subscribers, which were previously set to expire on July 12. These benefits will now remain in effect through July 19. Fable 5 usage continues to count against up to half of a subscriber's weekly limit at no extra cost.
- • Promotional access and 50% higher weekly rate limits for Claude Code are extended through July 19.
- • The extension follows the initial promotional window that was scheduled to end on July 12.
- • Fable 5 usage continues to count against up to half of a subscriber's weekly limit at no extra cost.
Developers have an additional week to utilize the higher rate limits and promotional access to Fable 5 for their coding and reasoning workflows.
11. Study Compares Token Efficiency of Claude Code and OpenCode
An empirical study was conducted to compare the efficiency of the agentic coding tools Claude Code and OpenCode. The study was initiated after anecdotal evidence suggested that Claude Code caused usage meters to rise more quickly than OpenCode. Researchers collected empirical data by adding logging between the coding tools and Anthropic's endpoint to capture all requests and usage blocks. The study concluded that Claude Code is less efficient than OpenCode in terms of its cache strategy and harness token usage, sending 33,000 tokens before reading the prompt compared to OpenCode's 7,000 tokens.
- • A study compared the efficiency of the agentic coding tools Claude Code and OpenCode by logging requests to Anthropic's endpoint.
- • The study found that Claude Code sends 33,000 tokens before reading the prompt, whereas OpenCode sends 7,000 tokens.
- • Researchers concluded that Claude Code is less efficient than OpenCode in its cache strategy and harness token usage.
- • The study was initiated after anecdotal evidence suggested Claude Code caused usage meters to rise more quickly than OpenCode.
Developers using Claude Code should be aware of its higher token overhead, which can accelerate API usage costs compared to alternative tools like OpenCode.
12. Benchmarking Parallel Agent Throughput on RTX 5090 with Qwen3.6 35B
A benchmark test using an RTX 5090 GPU and the Qwen 3.6 35B model loaded via LM Studio evaluated multi-agent throughput. The test parameters included 5 requests per agent, 1024 max tokens, and a temperature of 0.3. Combined throughput scales sub-linearly, with 8 agents providing 2.2 times the throughput of a single agent. Individual agent speed drops from 256.54 t/s with one agent to 67.22 t/s with eight agents. Peak efficiency of 70.4% is achieved when running 2 agents. The primary performance bottlenecks are identified as KV cache requirements and compute splitting.
- • A benchmark of Qwen3.6 35B loaded via LM Studio on an RTX 5090 evaluated multi-agent throughput with parallel tasks set to 8.
- • Combined throughput scales sub-linearly, with 8 agents providing 2.2 times the throughput of a single agent.
- • Individual agent speed drops from 256.54 t/s with one agent to 67.22 t/s with eight agents.
- • Peak efficiency of 70.4% is achieved when running 2 agents, while the recommended configuration for OpenCode is 4 to 5 agents.
- • Each agent utilizes a full context window, resulting in 8x KV cache usage when running 8 agents.
- • The primary performance bottlenecks are identified as KV cache requirements and compute splitting.
It provides concrete configuration guidelines for developers to balance throughput gains against VRAM and overhead costs when running local multi-agent systems.
13. Community Workarounds Address Qwen3.6-27B Tool-Calling and Looping Issues
Users of the Qwen3.6-27B local model have reported frequent tool-call failures and looping behavior that hinder its use as a reliable local alternative to frontier models. To address this, one user developed an extension for the Pi coding agent that monitors JSON streams for looping and automatically injects prompts to keep the model on track when it stops prematurely. Additionally, community members have identified that using specific chat templates, such as those hosted on Hugging Face by 'froggeric', can resolve common looping and tool-call failure issues in local models.
- • Users of Qwen3.6-27B reported frequent tool-call failures and looping behavior during local agent tasks.
- • One developer created an extension for the Pi coding agent that monitors JSON streams for looping and injects prompts to keep the model on track.
- • Community members identified that using specific chat templates, such as those hosted on Hugging Face by 'froggeric', can resolve common looping and tool-call failures.
These workarounds help developers run Qwen3.6-27B as a more reliable local alternative to frontier models for agentic tasks.