What does Vicki Boykis argue about running local AI models?

Boykis argues that open-weights local models have finally crossed the practical bar for real coding work. Running her own Pi agent harness against models like Gemma 4, she estimates they reach about 75% of the accuracy and speed of frontier cloud models on her agentic coding tasks, all without API dependencies. She frames this as a personal experiment, not a benchmark.

Which local models and tools does Vicki Boykis use?

Boykis runs open-weights models including Mistral 7B, Gemma 3, GPT-OSS-20B, several Qwen variants, Gemma-4-26b-a4b, and gemma-4-12b-qat. She runs them on a 2022 M2 Mac with 64GB RAM through llama.cpp, Ollama, LM Studio, Open WebUI, and Docker, using her own Pi agent harness for agentic coding loops.

How accurate are local models compared to frontier cloud models, according to Boykis?

Boykis estimates local models reach about 75% of the accuracy and speed of frontier cloud models on her agentic coding tasks. She is clear that she has no concrete scientific evidence for the figure and presents it as a personal experiment rather than a formal benchmark, noting that local models still face slow inference and limited context windows.

Why does this essay matter for developers?

The essay matters because it gives a new short answer to whether local models are worth using yet. Boykis reports the practical answer flipped from 'not really' to 'yes, with some patience,' which appeals to developers who care about privacy, offline work, or avoiding API bills. The post reached 653 points on Hacker News and reflects recent open-weights releases like Gemma 4.

Vicki Boykis · 2026-06-15 · notable

Vicki Boykis: 'Running Local Models Is Good Now'

ML engineer Vicki Boykis argues local models finally cleared the practical bar — she runs agentic coding workflows on Gemma 4 and friends, hitting about 75% of frontier cloud accuracy without API dependencies.

Cover image for Vicki Boykis' essay on running local language models.

A working ML engineer says open-weights local models are finally usable for real coding work.

Quick facts

Author	Vicki Boykis, ML engineer
Hardware	2022 M2 Mac, 64GB RAM
Accuracy estimate	~75% of frontier cloud models on agentic coding
Models run	Gemma 3/4, Mistral 7B, GPT-OSS-20B, several Qwen variants
Runtimes	llama.cpp, Ollama, LM Studio, Open WebUI, Docker
Hacker News	653 points

What is it?

An essay from Vicki Boykis, a founding ML engineer who builds recommendation systems and information retrieval at a startup and previously worked at Mozilla.ai. She walks through her current local-model setup and argues the on-device side has crossed a usability line in the last year.

How does it work?

Boykis runs her own 'Pi agent harness' against open-weights models — Mistral 7B, Gemma 3, GPT-OSS-20B, several Qwen variants, Gemma-4-26b-a4b, and gemma-4-12b-qat — through llama.cpp, Ollama, LM Studio, Open WebUI, and Docker. She estimates the models reach about 75% of the accuracy and speed of frontier cloud models on her agentic coding tasks. She notes she has 'no concrete scientific evidence' and frames this as a personal experiment, not a benchmark.

Why does it matter?

The post hit 653 points on Hacker News and is being shared as the new short answer to 'should I bother with local models yet.' For developers who care about privacy, offline work, or AI without API bills, the practical answer flipped from 'not really' to 'yes, with some patience.' It also confirms the impact of recent open-weights releases like Gemma 4.

Who is it for?

Indie developers, privacy-conscious teams, ML hobbyists

Frequently asked questions

What does Vicki Boykis argue about running local AI models?: Boykis argues that open-weights local models have finally crossed the practical bar for real coding work. Running her own Pi agent harness against models like Gemma 4, she estimates they reach about 75% of the accuracy and speed of frontier cloud models on her agentic coding tasks, all without API dependencies. She frames this as a personal experiment, not a benchmark.
Which local models and tools does Vicki Boykis use?: Boykis runs open-weights models including Mistral 7B, Gemma 3, GPT-OSS-20B, several Qwen variants, Gemma-4-26b-a4b, and gemma-4-12b-qat. She runs them on a 2022 M2 Mac with 64GB RAM through llama.cpp, Ollama, LM Studio, Open WebUI, and Docker, using her own Pi agent harness for agentic coding loops.
How accurate are local models compared to frontier cloud models, according to Boykis?: Boykis estimates local models reach about 75% of the accuracy and speed of frontier cloud models on her agentic coding tasks. She is clear that she has no concrete scientific evidence for the figure and presents it as a personal experiment rather than a formal benchmark, noting that local models still face slow inference and limited context windows.
Why does this essay matter for developers?: The essay matters because it gives a new short answer to whether local models are worth using yet. Boykis reports the practical answer flipped from 'not really' to 'yes, with some patience,' which appeals to developers who care about privacy, offline work, or avoiding API bills. The post reached 653 points on Hacker News and reflects recent open-weights releases like Gemma 4.