Commit

Quick recap for new folks: I built a small Neovim module that makes AI autocomplete ask the language server first. Before every suggestion, the model gets real facts from gopls / rust-analyzer: the signature of the call I'm in, the type before the dot, and what actually exists after it in the library version I have installed. Fewer invented methods, more code that compiles. Because it plugs into Minuet, I can swap the model with one line of config. So I spent a few days writing real Go with different models. Not a benchmark, just honest impressions. Prices are per 1M tokens, input/output. Open models: Nemotron Lightning 3.5 30B A3B (Fireworks), $0.05/$0.30 Cheap and fast, but inaccurate. Confident, dumb suggestions, like a "get rich" mentor on YouTube. DeepSeek V4.1 Flash (Fireworks), $0.30/$1.20 Cheap, on the slower side, and disappointing for code suggestions. Codestral (Mistral), $0.30/$0.90 Fast and cheap, but most of the time it returned nothing at all. When it did talk, it reminded me of Mr Freeze in Batman & Robin: full of bad one-liners. Qwen 3.8 27B (Cerebras), $0.99/$1.50 Pricey, super fast, and far better than the other open models. I'd happily use this combo, but the costs stack up. Closed models (where your requests may be logged): Claude Haiku 5.5, $0.10/$0.50 On the slower side and cheap. It likes to add chatty text around the code, but some suggestions feel like magic. Gemini 3.1 Flash-Lite, $0.25/$1.25 Super fast and very good at correct code. With the language-server context on top, it's on another level. If I could, I would. All day, all night. My takeaway: the model matters, but grounding matters just as much. Even small, fast models get noticeably better once they know the real API instead of guessing from training data. #Neovim #Golang #AI #DeveloperTools