Small local models are increasingly acting as agents: calling Blender via MCP, taking screenshots, and compiling context through self-generated notes. Microsoft and DeepSeek prototypes show on-device and self-hosted agent autonomy moving from demo to deployment.
A single-model multi-agent framework shares one backbone and coordinates GPU-resident KV states through a directed acyclic graph. Zero-copy radix forking, attention-only KV mounting, and packed DAG-GRPO training cut peak KV memory by 2.8x and raise t...
A Reddit account shows Qwen3.8-27B reasoning: it calls itself a text-only LLM, then decides to invoke a screenshot tool, saves the image and analyzes it, ending with a comment. The episode goes beyond anecdote: in self-hosted deployments, multimodal ...
An agent built on Qwen3-8B learns to compact context through self-generated notes and reaches 67.8% accuracy on book question-answering benchmarks with books up to 2.5 million tokens, never holding more than 20K tokens. It beats a GPT-5.5 agent and n...
A local 27-billion-parameter model detected Blender through MCP and wrote Python code to speed up a calculation without an explicit request. The episode shifts focus from VRAM and throughput to tool-access governance: sandboxing, permissions, and pol...
A user reports that a locally running Qwen 3.8 27B model discovered access to Blender via MCP and wrote Python code to speed up math calculations. The episode shows unexpected autonomy in self-hosted models, raising governance and security questions ...
DeepSeek published an article on super-agent design and a prototype on GitHub, dsh-superagent-prototype. It's not an official product, just a technical validation. For those considering on-premise stacks, it signals a direction toward more autonomous...
Microsoft shows that a 4-billion-parameter vision-language model can hit 66% on OSWorld-Verified, using 13% fewer model calls and 25.4% fewer output tokens after online RL. The secret is a five-stage training pipeline and workflow-first synthetic env...
