Show HN: Supermaven, the first code completion tool with 300k token context
A new cloud-based code completion tool, Supermaven, promises extremely fast suggestions and a 300,000‑token context window, leading some developers to report noticeably better results than GitHub Copilot, especially on large codebases. Others raise concerns about its requirement for a credit card to start a trial, limited transparency around data usage and training sources, and the legal risks of generating code potentially influenced by GPL or other restrictive licenses. The conversation also highlights broader unease about AI-assisted coding: its impact on developer focus and learning, over-reliance by less experienced engineers, and the trade-off between productivity gains and code quality or privacy.
Overall reception & performance
- Several users report Supermaven feels faster and often better than Copilot, with more grounded, less “hallucinated” suggestions, especially across a codebase.
- Others find the completion UX distracting, as very frequent suggestions pull them out of “writing mode” into “evaluation mode.”
Model architecture & long context
- The project claims a new architecture more efficient than Transformers for long context; commenters speculate it could be SSM-based (e.g., Mamba-like) or similar to other long-context methods, but details are proprietary.
- Some compare the 300k context claim to Gemini’s long contexts and wonder how it really works without RAG.
- A few users want demonstrations of genuinely long-range, cross-file reasoning, not just fast per-keystroke completion.
Data, privacy & training sources
- Users ask what code the model is trained on and whether software licenses (including GPL/AGPL) are respected; no concrete answer appears in the thread.
- Questions are raised about whether customer source code is used for training or fine-tuning; this is also left unclear.
- The tool uploads the entire git repo to the cloud and retains data up to 7 days; some criticize the privacy policy as vague and request clearer documentation.
- There is broader debate on whether LLM training on copyrighted code is analogous to human reading, and whether copyright law will treat them similarly, with jurisdictional differences (e.g., Japan) noted.
Onboarding, pricing & trial
- Many criticize the requirement of a credit card for the 30‑day free trial, calling it high-friction and potentially predatory if it relies on users forgetting to cancel.
- Others defend it as necessary due to AI inference costs, though some suggest shorter trials or non-auto-renewing “credit-based” trials.
- Some note this conflicts with “Show HN” norms of low-barrier trying.
Tooling, UX & ecosystem
- All processing is cloud-based; no local or hybrid mode currently.
- Users request IntelliJ, Neovim, and more general LSP-style integration; an IntelliJ plugin is promised soon.
- Some prefer chat or edit-prompt workflows (like Cursor/Zed) over inline completion and ask for such modes.
- Questions arise about support for non-mainstream languages; coverage is not clearly specified.
Developer impact & code quality
- Supporters emphasize speed, reduced boilerplate, and RSI prevention.
- Skeptics worry about novices blindly accepting code, worsening overall software quality and weakening fundamental skills.
- Several argue these tools are most beneficial in the hands of experienced developers who can review and direct the output.
Stability & security concerns
- A few users report Windows issues (agent window popping up, unclear licensing prompts).
- One VirusTotal scan shows detections from two engines; it’s unclear whether this is a false positive.