1. TensorSharp Adds Local Support for Qwen-Image 2.1 and LoRA Adapters
TensorSharp has updated its platform to support the Qwen-Image 2.1 model, which was released by Alibaba on September 20. This update enables local text-to-image generation and editing, including support for Qwen-Image 2.1 LoRA adapters such as Pruna 8-step and Viggle Turbo. Developers can now run the model via the TensorSharp CLI by specifying a base configuration file and optional LoRA settings.
- • TensorSharp now supports local execution of the Qwen-Image 2.1 model released on September 20.
- • The update adds support for Qwen-Image 2.1 LoRA adapters, including Pruna 8-step and Viggle Turbo.
- • Developers can access multiple model files, including Qwen-Image 2.1 GGUF and Qwen-Image 2.1 VAE.
- • The engine is accessible via the TensorSharp CLI using a base configuration file and optional LoRA settings.
This integration allows developers to run the recently released Qwen-Image 2.1 model locally with custom LoRA adapters for specialized image generation and editing tasks.
2. KoboldCpp Introduces Built-In Agent Harness and Warns of Phishing Site
KoboldCpp has introduced a built-in Agent Harness designed as a lightweight alternative to tools like Claude Code. The harness includes 9 built-in tools, supports a 2k token system prompt, and allows developers to expand functionality by loading an mcp.json file to integrate MCP tools. Alongside this release, the developer issued an urgent warning regarding a phishing site, kobolcpp.com, which uses blackhat SEO to distribute malware.
- • KoboldCpp has added an integrated Agent Harness featuring 9 built-in tools and a 2k token system prompt.
- • The agent runtime supports Model Context Protocol (MCP) tools by loading an mcp.json configuration file.
- • The agent can be enabled via the GUI launcher's Admin tab or the --agent command-line flag, and supports OpenAI-compatible endpoints.
- • Recommended system requirements for the agent include at least 28k context, 8k generation limit, and 12GB of VRAM.
- • The developer issued an urgent warning against the phishing site kobolcpp.com, which uses blackhat SEO to distribute malware.
It gives developers a lightweight, local agent runtime with built-in tools and MCP support directly inside KoboldCpp, while warning them to avoid a malware-distributing phishing site.
3. DeepSeek Introduces DSec Production Sandbox Platform for Agents
DeepSeek Elastic Compute (DSec) is a production-scale sandbox platform designed for large-scale agentic training and evaluation. The platform provides a unified SDK supporting function calling, container, microVM, and full-VM sandbox backends. DSec utilizes the Fire-Flyer File System (3FS) to load image data on demand across clusters and is co-designed with a reinforcement learning framework to decouple stateful rollout execution from preemptible GPU training.
- • DeepSeek Elastic Compute (DSec) is a production-grade sandbox platform designed for large-scale agentic training and evaluation.
- • The platform offers a unified SDK supporting multiple backends, including function calling, containers, microVMs, and full virtual machines.
- • DSec utilizes the Fire-Flyer File System (3FS) to load container image data on demand across clusters.
- • A single production-scale unit of DSec spans roughly 160 nodes, supporting over 380,000 concurrent sandboxes and 3 million daily runs.
- • The system is co-designed with a reinforcement learning framework to isolate stateful agent execution from GPU training resources.
It provides a highly scalable, unified sandbox infrastructure for running, training, and evaluating agentic workflows safely at production scale.
4. Drawgent Integrates Coding Agents with Live Excalidraw Whiteboards
Drawgent is a coding agent tool that integrates Claude Code, Codex, or opencode with an Excalidraw whiteboard for live diagram editing. The tool uses the Agent Communication Protocol (ACP) to interface with coding agents and provides a dedicated suite of MCP tools, including get_scene, add_elements, and update_elements. Drawgent supports headless Chrome or Chromium for rendering and allows agents to join collaborative Excalidraw rooms as active collaborators.
- • Drawgent connects coding agents like Claude Code or Codex to an Excalidraw canvas for live, interactive diagram editing.
- • The tool provides a dedicated suite of MCP tools, including get_scene, add_elements, and update_elements, for agentic canvas manipulation.
- • Agents can join collaborative Excalidraw rooms as active participants, responding to canvas text triggers or chat panel commands.
- • The system runs on Node.js (version 18+) and uses headless Chrome or Chromium to render the Excalidraw canvas.
- • Scenes are saved locally in a git-ignored JSON file, and the tool exposes REST and WebSocket APIs for external integrations.
It enables developers to integrate coding agents directly into collaborative whiteboards, allowing agents to read, edit, and verify diagrams in real time.
5. Reladraw Diagramming Language Offers AI-Friendly Layout Control
Reladraw is a new diagramming language designed to balance the automatic placement capabilities of tools like Mermaid with the manual control of software like Draw.io. The tool is optimized for easy manipulation by both humans and AI agents. It is installable via npm and provides a pre-built skill for direct integration with Claude and other AI agents, allowing developers to easily generate structured visual layouts.
- • Reladraw is a diagramming language designed to balance automatic layout placement with precise manual control.
- • The tool is optimized for easy manipulation by both human developers and AI agents.
- • It is installable via npm and includes a pre-built skill for direct integration with Claude and other agents.
- • A web-based playground is available on GitHub for testing without local installation.
It provides an AI-friendly diagramming language that developers can easily integrate into Claude and other agents to generate structured, controllable visual layouts.
6. Splash 1.1.0 Released with GGUF and MLX Support for Apple Silicon
Splash version 1.1.0 has been released, offering an optimized local inference runtime for Apple Silicon. The software combines optimized kernels, speculative decoding, prefix caching, and mixed-weight support in a single program. In community testing, a user reported running the Qwen3.8 27B model in Unsloth's UD-Q4_K_XL format on an M5 Pro with 64GB of RAM, achieving speeds of 50 tokens per second.
- • Splash version 1.1.0 has been released, bringing GGUF quantization support and MLX imports to Apple Silicon.
- • The software combines optimized kernels, speculative decoding, prefix caching, and mixed-weight support in a single runtime.
- • A user reported achieving 50 tokens per second running the Qwen3.8 27B model in Unsloth's UD-Q4_K_XL format on an M5 Pro with 64GB of RAM.
It gives Apple Silicon developers a highly optimized local inference engine that supports GGUF, MLX imports, and speculative decoding for fast model execution.
7. Fixed GPT-OSS Jinja Template Resolves Dropped Answers in Chat History
A bug was identified in the GPT-OSS Jinja template (originally based on Unsloth's version) that causes the model's answer content to be dropped when chat history containing reasoning is replayed. This bug caused GPT-OSS 20B to malfunction, while the larger 120B variant remained functional. A fixed template has been released on Hugging Face, which also introduces a preserve_thinking feature that improves multi-turn inference speed through prefix caching.
- • A bug in the GPT-OSS Jinja template (derived from Unsloth) was found to drop model answers when replaying chat histories containing reasoning.
- • The bug caused the GPT-OSS 20B model to malfunction, while the larger 120B variant remained functional.
- • A fixed template has been released on Hugging Face, resolving the copy-paste error from the reference template.
- • The updated template introduces a preserve_thinking feature, which uses prefix caching to accelerate multi-turn inference at the cost of higher token usage.
It fixes a critical bug in GPT-OSS templates that causes reasoning-heavy chat histories to drop answers, while introducing a prefix-caching feature to speed up multi-turn inference.
8. Prompting Method Enables Single-Token Decisions in GLM-5.3-Flash
Developers can now configure standard LLMs to achieve Jev-like decision properties by crafting prompts where the first output token answers the question. This method enables decision-making within a single forward pass and has been documented for use with GLM-5.3-Flash and the vLLM inference engine. The setup supports vision inputs, making it highly applicable for fast visual classification tasks.
- • A new prompting method enables standard LLMs to make decisions within a single forward pass by forcing the first output token to answer the question.
- • The approach is documented and tested using GLM-5.3-Flash and the vLLM inference engine.
- • The setup supports vision inputs, allowing for fast visual classification tasks.
- • Benchmarks indicate the setup matches Jev in accuracy and speed, outperforming Laya, though Jev remains more cost-effective.
It allows developers to achieve ultra-low latency decision-making in standard LLMs like GLM-5.3-Flash using a single forward pass.
9. Single-Function Wrapper Enables Fast Vision Classification via LLM Logprobs
A lightweight technique allows developers to perform computer vision tasks by reading an LLM's token probabilities instead of generating full text. The method requires setting max_completion_tokens to 1, logprobs to true, and top_logprobs to 20 in a standard Chat Completions request. The author extended the Jev project's request format to support base64-encoded image data in an attachments field, providing a Python script that uses OpenCV for webcam frame capture.
- • The wrapper performs computer vision classification by reading an LLM's token probabilities for specific questions.
- • The technique requires configuring Chat Completions requests with max_completion_tokens set to 1, logprobs set to true, and top_logprobs set to 20.
- • The implementation extends the Jev project's format to support base64-encoded image data in an attachments field.
- • A provided Python script uses OpenCV for webcam capture and is compatible with both local llama.cpp and OpenAI API endpoints.
- • Performance reached approximately 1 frame per second using Gemma 4 12B on an RTX 3090, and 0.2 frames per second with OpenAI's gpt-6-luna.
It provides a lightweight, single-function wrapper to perform fast computer vision classification by reading LLM token probabilities instead of generating full text.
10. FreeToken Framework Adds Overspill Disk Tier for Massive MoE Models
Building on the FreeToken framework's existing capabilities for local MoE serving, the new experimental Overspill disk tier allows models to exceed available system RAM. In tests, an 85 GB DeepSeek-V4-Flash REAP-150B model was successfully run on an RTX 3060 with 64 GB of DDR5 RAM, achieving 2.8 to 3.4 tokens per second. The implementation uses memory-mapping for experts and parallel reads to optimize performance.
- • Overspill is a new experimental disk tier for the FreeToken framework.
- • It enables running MoE models that exceed available system RAM.
- • Tests achieved 2.8 to 3.4 tokens per second for an 85 GB model on an RTX 3060.
- • The framework uses memory-mapping and parallel reads to manage expert loading.
- • The project is open-source under the Apache-2.0 license.
This development expands the FreeToken framework's utility by enabling the execution of massive MoE models on hardware with limited RAM, providing a new option for local model serving.