2026-10-15
Local AI agents with open-weights models
Kimi K3, DeepSeek-V4.1-Flash, MiMo-V2.6 and GLM-5.3 rival the closed models on tool use, automation and computer use. Locally the agent works on your data without leaking it, and you pay by the hour, not the token.
An agent doesn't answer a question: it plans, calls tools, reads the results and retries. A single task can run for hours and touch the CRM, mail, internal documents and databases. It's the use case where local execution matters most, because every tool call carries company data with it.
The strongest models on agents
Kimi K3 is the reference for research and tool use. In Moonshot AI's table it scores 91.2 on BrowseComp (GPT-5.6 Sol 90.4, Claude Fable 5 88.0), 94.5 on MCPMark-Verified and 84.8 on OSWorld-Verified. It's built for long sessions with minimal oversight, on contexts up to a million tokens. In the catalog it runs with vLLM on 8× B300 at about 215 tok/s.
DeepSeek-V4.1-Flash is designed for input-heavy agentic workloads: it activates only 8B params in prefill, and the 890-byte-per-token KV cache makes long contexts full of tool output cheap. DeepSeek reports 54.8 on AutomationBench (Claude Opus 5.0 50.3) and 31.8 on Agent's Last Exam, the highest in its table. You start it with vLLM on 8× H200 from about $16.5/h, or locally on a 128 GB Mac with ds4 Metal.
MiMo-V2.6 from Xiaomi is omnimodal (text, image, video, audio) with a single reinforcement-learning run mixing coding, agents, vision and cybersecurity, so the skills reinforce each other. The Flash-RL variant scores 73.6 on Toolathlon-Verified and 52.3 on AutomationBench, and runs on 4× RTX PRO 6000 at about $3.1/h: the cheapest agentic path in the catalog.
GLM-5.3 is strong on long-horizon tasks: 48.2 on AutomationBench and 73.0 on Toolathlon-Verified in Z.ai's table, with context up to 1M tokens.
Scores are vendor-reported, with different harnesses from one card to the next: use them to orient, not as an absolute ranking.
Why keep them in-house
- Data doesn't leave. The agent reads and writes to your systems, but the model answers on an endpoint exposed only on
127.0.0.1. Nothing goes to a third-party API. - Cost is by the hour, not the token. An agent running for hours burns millions of tokens. With a dedicated GPU the bill is predictable and the session stops when you're done.
- Scoped, expiring MCP keys. SophyAI issues limited MCP keys to give agents only the tools they need, for as long as they need them.
How it connects
SophyAI publishes a local OpenAI-compatible endpoint: the agentic frameworks in the catalog point at it — Agno, LangGraph, CrewAI, smolagents, OpenHands, Letta, n8n — with no code change. Pick a recipe, start the session (local or cloud) and the agent talks to a model running under your control.
For the other two scenarios see agentic coding, running locally and cybersecurity with local models. The full model overview is in Frontier open-weights models, running locally.