CloudSurf

Everyone has the power to build and deploy reliable apps, from any device, for all platforms.

CloudSurf Software LLC builds CloudSurf: a workspace where documents, websites and apps are one plain-text format, SurfDoc, and an assistant, Surfy, writes and edits it with you. You describe what you want from your phone; CloudSurf builds it, deploys it and keeps it reliable. The same workspace serves the professional developer and the person who has never written code, because both work on the same file.

• The web app at app.cloudsurf.com and the Mac app from cloudsurf.com; the iPhone and Android apps are in testing.
• The format's parser, surf-parse, is open source under MIT and on crates.io.

We also train our own models so that CloudSurf can run on your device, with no network and no per-token cost: small, US-trained, open-weight models released under Apache-2.0.

Models

• CloudSurf-1B · GGUF — a 0.89B-parameter base model trained from random weights on 18.16B tokens: our tokenizer, our trainer, our corpus pipeline, our weights. It is the on-device tier of Surfy: routing requests to tools and writing SurfDoc on a phone or a laptop. Every bar was stamped before training and the verdict is on the card whichever way it landed, including the one it failed.
• CloudSurf-4B-FC · GGUF — Gemma-4 E4B (4B active / ~8.0B total params) fine-tuned for function calling and agentic tool use. BFCL V4 FULL, official harness, 3-seed mean: 55.73 vs 34.81 for the stock base, above the published small-model class bar (51.40). Self-run numbers with full settings disclosure; leaderboard submission open at gorilla PR #1357.

Evidence, not just claims

Every score we publish ships with the raw artifacts to check it:

• CloudSurf-4B-FC-bfcl-results — raw BFCL V4 result files for our runs and the stock baselines, with per-category scores.
• tau2-trajectories-cloudsurf-4b-fc — full τ²-bench evaluation trajectories backing our leaderboard submission, open at tau2-bench PR #480.

Model cards disclose both comparison frames, measured eval-noise bands, pre-registered bars and known limitations. Nothing written by a language model goes into our pretraining corpora, and the evaluation holdouts are never trained on.

Links

• App: app.cloudsurf.com
• Company: cloudsurf.com
• Code: github.com/cloudsurf-software