A 64GB Mac, and the local model that stopped needing a check
Vicki Boykis ran agentic coding on a 2022 M2 Mac with 64GB of RAM, and one open-weight model finally crossed a line: she stopped double-checking it against an API.
Her quality test isn't a benchmark. It's whether she has to verify the output against a paid model — and OpenAI's OSS-20B was the first where she did that a lot less often.
The thread's counter is blunt: boot Gemma 4 MoE expecting Claude Code behavior and you'll be very disappointed. Capable, but still far behind the frontier.
The day a free model stopped needing a chaperone
For two years the honest answer about local LLMs was "good enough to demo, not to ship." Vicki Boykis's post argues that quietly flipped, and it landed on Hacker News at 1028 points because she names the exact moment it happened for her.
The setup is deliberately unexotic: a 2022 M2 Mac, 64GB of RAM, 1TB of storage. On that, agentic coding workflows peak around 64GB of K-V cache. She tested Mistral 7B, Gemma 3 and 4, Qwen, and OpenAI's OSS-20B, and reaches roughly 75% of frontier accuracy for agentic coding. Her measure of "good enough" is not a leaderboard. It's a personal vibe metric: "do I have to double-check it against an API model." OSS-20B was the first where she stopped doing that so often. The deeper question she raises: "if we are constrained by performance and price, what architectural tradeoffs do we need to make?"
The 30B mixture-of-experts sweet spot
The sharpest practitioner note in the thread points at a specific class of model, not a vibe.
"there's something very interesting about a local model that does mixture-of-experts … and has in the order of 30 billion parameters … use around 20-30GB of RAM" — Simon Willison, creator of Datasette
Another commenter clocked real throughput: an 8-bit Qwen3.6 27B at 30 tok/s on llama.cpp, landing at "the level of SotA mid-late 2025."
"its 8 bit version runs at 30tok/s … it gives more-or-less the level of SotA mid-late 2025" — a developer benchmarking local LLMs on Apple Silicon
The asterisk nobody on the post wants printed
The enthusiasm has a hard floor. One owner of a 128GB rig who wants local to win still warned against the genre of post claiming parity.
"You can be far more ambiguous with your tasks with the larger proprietary models … with local models you need to be much more detailed in your prompt" — a user running both 128GB-class local and the big proprietary models
The rule that changed isn't "local replaces the frontier." It's that the gap is now a workflow tax, not a wall: a frontier model writes the plan, a 30B local model executes it, and you trade dollars for tokens and prompt discipline. For the labs renting access by the month, that arithmetic is the threat — customers can now multiply the subscription by 24 and ask whether hardware pays for itself first.
Get breakdowns like this every day
The top Hacker News stories, unpacked — by email, Telegram, or RSS.
Chrome 150 pulls the last MV2 flag on June 30, ending uBlock Origin
Google is removing the final escape hatch for Manifest V2 extensions. uBlock Origin is the last big ad blocker still on V2 — and after Chrome 150, the workaround flag is gone.