← Capa | Mercado

Local LLMs hit 90% on agentic coding eval as Ollama wins developer mindshare

2026-11-04 · Fonte:InfoWorld

Local LLMs hit 90% on agentic coding eval as Ollama wins developer mindshare

Open-weight models on laptops now score 90% on an agentic coding benchmark that floored local tools months ago, and Ollama has become a popular way to run them.

Developers who wrote off locally hosted large language models after early disappointments may want to take another look, according to an InfoWorld analysis of the Ollama runtime. The turnaround is stark: Simon P. Couch, senior software engineer at Posit, noted on Bluesky this spring that a few months earlier, every LLM he could run on his MacBook scored 0% on an agentic coding evaluation he assembled. The April releases of Qwen 3.5 and Gemma 4, he reported, both reached 90% on the same test.

Open-source developer Simon Willison, who tracks the AI industry closely, struck a similar note in a lightning talk at PyCon US 2026 in May, observing that models available on laptops, though far weaker than the frontier, have begun to wildly outperform expectations. The caveat stands: a model running on a laptop still cannot approach what state-of-the-art systems from Anthropic or OpenAI deliver in the cloud. For well-defined work — answering coding questions, writing functions, or summarizing documents — however, local models can be surprisingly capable.

Among the many routes to running models locally, Ollama stands out for popularity and easy setup, even if it is not the fastest option available. Its integration footprint is broad: mainstream programming tools including Visual Studio Code, JetBrains AI Assistant, Zed, and Posit Assistant support it out of the box. Ollama can also launch Claude Code or Codex with the option of pointing those coding agents at a local LLM.

Ollama is not the only choice. Alternatives for running local models on a PC or Mac include LM Studio, Jan, Unsloth, the LLM tool created by Simon Willison, and llama.cpp. For organizations weighing whether on-device inference can absorb routine coding and summarization workloads, the practitioner consensus is that the hardware already on the desk may now clear a bar it failed just months earlier.

Tags: Ollama · Simon P. Couch · Simon Willison · Local LLMs · Qwen 3.5 · Gemma 4


Leituras relacionadas


萬安算交所 AIXX · AIXX · Exchange verificado · Ir ao mercado