2026-10-15
Agentic coding, running locally
On SWE-bench, DeepSWE and Terminal-Bench the open-weights models are now level with the best closed ones. With SophyAI you connect them to Cursor, OpenCode, Aider, Continue and Cline via a local endpoint, without sending your code out.
Agentic coding isn't autocomplete: the model reads the repository, opens files, runs terminal commands, executes tests and iterates. It's the toughest test, and also the one where the open-weights models have closed the gap.
The numbers reported by the vendors
| Model | DeepSWE v1.1 | Terminal-Bench 2.1 | Notes |
|---|---|---|---|
| DeepSeek-V4.1-Flash | 74.2 | 90.6 | Codeforces 3471, the highest in its table |
| Kimi K3 | 67.5 | 88.3 | FrontierSWE 81.2; ProgramBench 77.8 |
| MiMo-V2.6-Pro | 71.9 | 89.9 | Terminal-Bench 4.0 34.9 |
| GLM-5.3 | 66.9 | 88.2 | Open SOTA on Terminal-Bench 3.0 (28.3) |
| GLM-5.3-Flash | — | — | 18B active, near Opus 4.8 at one-tenth the cost |
| Qwen3.8-Flash-Next | 58.7 | — | SWE-bench Pro 62.5; SWE-bench Multilingual 81.0 |
For reference, in the same tables the best closed models sit around 73–74 on DeepSWE v1.1 (Claude Opus 74.0, GPT-5.6 Sol 73.0). Scores are vendor-reported with different harnesses and settings, so they aren't one-to-one comparable — but they show how far the gap has shrunk.
Which recipe to pick
- One GPU, low cost. Qwen3.8-Flash-Next in GGUF on 1× RTX PRO 6000 does ~95 tok/s at ~$1.7/h: light enough to keep running all working day. GLM-5.3-Flash on the same GPU does ~51 tok/s at ~$1.7/h.
- Top code quality. GLM-5.3 and DeepSeek-V4.1-Flash are the strongest open models here. DeepSeek-V4.1-Flash also runs locally on a 128 GB Mac with ds4 Metal in Q2, or with vLLM on 8× H200 from ~$16.5/h for long contexts and speed.
- Long horizons. All of these reach 1M tokens of context: a whole repository fits without pruning.
How you connect it to your editor
SophyAI publishes an OpenAI endpoint on 127.0.0.1. You set it as the base URL in Cursor, OpenCode, Aider, Continue or Cline and the assistant uses the local model. Customer code never leaves your network, and the cost is the GPU hour, not the token.
The other two scenarios: local AI agents and cybersecurity with local models. Full overview: Frontier open-weights models, running locally.