Frontier open-weights models, running locally

By the autumn of 2026 the gap between closed and open-weights models has nearly closed. On the hardest coding, agentic and security benchmarks (Terminal-Bench, DeepSWE, CyberGym) the models you download from Hugging Face sit a few points away from GPT-5.6 Sol, Claude Opus and Claude Fable 5 — and beat them on some.

For a company the question changes: not which API to use, but where to run the model. With SophyAI the answer is on your Mac, or on a cloud GPU you start yourself that answers on an endpoint exposed only locally. Prompts, code and documents never pass through a third-party API.

The frontier models in the catalog

Parameters, context and licenses come from the Hugging Face model cards (October 2026). Speeds and hourly costs are from the SophyAI catalog: measured by us on Vast.ai where noted, otherwise the reference offer price. The app always shows the real price before you start.

What the model cards say

• Kimi K3 from Moonshot AI is the first open 3T-class model: 896 experts, 16 active per token, native vision and a one-million-token context. On agentic benchmarks it leads: 91.2 on BrowseComp and 94.5 on MCPMark-Verified, ahead of GPT-5.6 Sol in the vendor's table.
• GLM-5.3 from Z.ai is billed as the strongest open-weights model for coding, with a 50% gain over GLM-5.2 on its own Code Bench. On CyberGym, which measures vulnerability discovery, it scores 84.5.
• GLM-5.3-Flash is the first natively multimodal GLM: 18B active, hybrid sparse-and-linear attention, MIT license. Z.ai describes it as approaching Claude Opus 4.8 on coding and agents at one-tenth the cost of GLM-5.2.
• DeepSeek-V4.1-Flash introduces a causal encoder-decoder architecture: it activates 8B params in prefill and 16B in decode, with an 890-byte-per-token KV cache, about a quarter of V4-Flash. In DeepSeek's table it scores 74.2 on DeepSWE v1.1 and 90.6 on Terminal-Bench 2.1.
• MiMo-V2.6 from Xiaomi is omnimodal (text, image, video, audio) and trained with a single reinforcement-learning run mixing coding, agents, vision and cybersecurity. The Flash version scores 95.1 on CyberGym.
• Hy4-preview from Tencent is a 770B MoE with 49B active, Apache 2.0. In a blind side-by-side on 203 engineering tasks, rated by 163 internal experts, it comes out slightly ahead of both GLM-5.3 and Kimi K3.
• Qwen3.8-Flash-Next is the preview of the Qwen4 architecture: only 6B active params, yet 62.5 on SWE-bench Pro and 81.0 on SWE-bench Multilingual. It runs on a single GPU.

Scores are vendor-reported, with different harnesses and settings from one card to the next: use them to orient, not as an absolute ranking.

Why run them locally

• Confidentiality: source code, security logs, contracts and customer data stay on machines you control. The SophyAI endpoint listens on 127.0.0.1.
• Predictable cost: an agent that works for hours consumes millions of tokens. With a dedicated GPU you pay by the hour, not the token, and you stop the session when you're done.
• No third-party filters: for a red team or a SOC, an API that refuses a legitimate exploit analysis is an operational blocker. Locally, you set the usage policy.
• Reproducibility: the weights don't shift under you. The version you validated is the one that runs.

Read more

We wrote three articles on the uses where these models pay off the most:

• Local AI agents with open-weights models
• Agentic coding, running locally
• Cybersecurity with local models

← All posts