Sage 1.5B

Sage predicts your next code edit. Give it the lines around your cursor and the last few changes you made, and it writes back that window with the edit finished: the rest of a line you are halfway through typing, or the next change a rename or a new parameter calls for.

It is small enough to run on a laptop with llama.cpp, and it speaks the Sweep next-edit prompt format, so Zed's edit predictions can use it with no plugin.

What it was trained on

Sage is a full fine-tune of sweepai/sweep-next-edit-1.5B (Apache-2.0, built on Qwen2.5-Coder-1.5B) for one epoch on 58M tokens:

• Edit histories. Every multi-hunk commit is replayed hunk by hunk: the earlier hunks become the recent changes, and the next hunk is the target. About 30,000 examples come from the recent history of 20 permissively licensed projects (React, TypeScript, Next.js, Vite, Deno, Tailwind CSS, FastAPI, Pydantic, Django, Transformers, Ruff, uv, Kubernetes, Prometheus, GitHub CLI, Tokio, Swift, Flutter, VS Code and Hermes Agent) and about 23,000 from the author's own projects.
• Accepted review suggestions. 1,248 GitHub suggestion comments from the same projects whose suggested code is in the merged result. The reviewer's note goes into the prompt ahead of the file.
• Half-typed code. Half of all edits are shown cut at a random character, with the cursor on the cut line, the way an editor sends them while you type. The target is the finished edit.

Examples containing anything that looks like a credential (known key formats and long high-entropy strings) were dropped, as were .env, key and lock files.

Prompt format

<|file_sep|>{path}.diff
original:
{lines before your last change}
updated:
{lines after it}
<|file_sep|>original/{path}
{the window before your last change}
<|file_sep|>current/{path}
{the window now: 10 lines above the cursor, the cursor line, 10 below}
<|file_sep|>updated/{path}

Up to three .diff blocks, oldest first. Sage answers with the updated window and stops at <|file_sep|> or end of text. Decode greedily.

Use it in Zed

Serve the GGUF with llama.cpp:

llama-server -m sage-1.5b-q8_0.gguf --alias sage --port 8080 -c 8192 -fa on -ngl 99 \
  --cache-reuse 256 --spec-type ngram-simple --draft-max 16

The output is mostly a copy of the current window, so n-gram speculation (drafting from the prompt itself) roughly triples decoding speed, and cache reuse keeps the prompt between keystrokes.

Then in Zed's settings.json:

{
  "edit_predictions": {
    "provider": "open_ai_compatible_api",
    "open_ai_compatible_api": {
      "api_url": "http://127.0.0.1:8080/v1/completions",
      "model": "sage",
      "prompt_format": "sweep",
      "max_output_tokens": 512
    }
  }
}

How well it does

Exact match of the whole predicted window, greedy, on held-out data (the most recent commits of each project, and review suggestions from pull requests not seen in training):

	Sweep 1.5B	Sage 1.5B
Half-typed code (300 examples)	–	32.7%
Next edit, the author's projects (300)	3.7%	17.3%
Next edit, the 20 public projects (300)	5.3%	29.0%
Accepted review suggestions (161)	0.0%	11.8%

Sweep's own numbers here use the windows this card describes, which are not the windows Sweep was trained on, so they understate Sweep in its own editor.

On an Apple M1 Mac with 8 GB of RAM, using the q8_0 GGUF and the settings above: about 1 second for a suggestion when the prompt is already cached (around 150 tokens/s, 85% of drafted tokens accepted), and 5 to 7 seconds for the first one in a new place, most of it spent swapping on a machine short of memory. More RAM should help most.

Limitations

• A bare keyword typed just above a line that starts with the same keyword can be taken for an accidental duplicate and removed. A few more characters fix it.
• Many real edits are new features that no model could guess from the previous edit; those are most of the misses.
• Part of the training data is the author's own code, so Sage has seen it and may echo patterns from it.
• It writes the whole 21-line window each time, so it is slower than models that emit only the changed lines.

Files

• model.safetensors and tokenizer files: the weights for transformers.
• sage-1.5b-q8_0.gguf: for llama.cpp and anything built on it.
• metrics.json: the evaluation behind the table above.

License

Apache-2.0, as is the base model. The training data comes from repositories under MIT, Apache-2.0 and BSD-3-Clause licences.