[SYSTEM: INITIALIZING]

REFRACTRPRIVATE AI INFRASTRUCTURE
PROJECT BYREFACTR
BOOT PIPELINEVERIFY. ISOLATE. RUN.
  1. EnvironmentWAIT
  2. Data boundaryWAIT
  3. Model runtimeWAIT
  4. Inference pathWAIT
refactr

Stop renting Intelligence, own it.

Private AI models and agents, deployed on your own servers. Your data stays yours.

Introducing Factr-I

Your AI agent, on your machine, learning how you work.

Factr-I is a lightweight desktop agent that keeps its own memory, learns from your work, and runs on the model you already have.

Your model can talk. Factr-I puts it to work.

Answers arrive with their source.

Ask a question about your files and Factr-I answers with the source passage open beside it, so you can check the answer before anyone acts on it.

Code arrives tested.

Describe the change and Factr-I edits the code, shows the diff and runs your tests before it hands the work back.

Work ends in a file.

Give it a task with several steps and it works through them on your machine, then returns the finished file instead of a chat reply.

Answers arrive with their source.

Ask a question about your files and Factr-I answers with the source passage open beside it, so you can check the answer before anyone acts on it.

Factr-I's response to a question about carry-over PTO, shown in full under the window title.
The Carryover section of the Paid Time Off Policy, open as the source of the answer.

Code arrives tested.

Describe the change and Factr-I edits the code, shows the diff and runs your tests before it hands the work back.

The request to add a health endpoint and Factr-I's response: it updated the endpoint and verified that all 9 tests pass.
The Changes panel listing the changed files with their added and removed lines.

Work ends in a file.

Give it a task with several steps and it works through them on your machine, then returns the finished file instead of a chat reply.

The task to build a comparison spreadsheet and Factr-I's response confirming that it created and verified the file.
The finished spreadsheet with a row each for Americas, APAC and EMEA.

Performance and Cost

Factr-I, Prime and Hermes ran the same 1,067 tasks from six public benchmarks, all on GPT-6 Luna.

Showing Score. Score: higher is better. Tokens and cost: lower is better.
Score. Score: higher is better. Tokens and cost: lower is better
BenchmarkFactr-I68.1M tokens$3.811Prime (Rust)2159.7M tokens$13.49Prime135.2M tokens$4.61Hermes103.8M tokens$3.24
Long-context pairsOOLONG-Pairs · 20 tasks73.8% (best)F123.4%F132.7%F129.8%F1
Long-context reasoningOOLONG · 50 tasks90.0% (best)45/5048.0%24/5060.0%30/5034.0%17/50
General assistantGAIA · 165 tasks40.0% (best)66/16539.4%65/16537.6%62/16528.5%47/165
Terminal tasksTBLite · 100 tasks68% (best)68/10059%59/10064%64/10060%60/100
Code editingPolyglot · 225 tasks86.2% (best)194/22570.2%158/22573.3%165/22581.8%184/225
Long reasoningLongCoT-Mini · 507 tasks19.3% (best)98/50719.3% (best)98/50717.6%89/50717.8%90/507
Memory. Median peak memory (MiB): lower is better. Tokens and cost: lower is better
BenchmarkFactr-I68.1M tokens$3.811Prime (Rust)2159.7M tokens$13.49Prime135.2M tokens$4.61Hermes103.8M tokens$3.24
Long-context pairsOOLONG-Pairs · 20 tasks84MiB72 (best)MiB627MiB231MiB
Long-context reasoningOOLONG · 50 tasks96MiB71 (best)MiB618MiB221MiB
General assistantGAIA · 165 tasks53 (best)MiB103MiB637MiB234MiB
Terminal tasksTBLite · 100 tasks27MiB46MiB778MiB241MiB
Code editingPolyglot · 225 tasks390MiB255 (best)MiB826MiB416MiB
Long reasoningLongCoT-Mini · 507 tasks68MiB60 (best)MiB609MiB226MiB

Read the head-to-head comparisons for one competitor at a time, including where each of them leads.

  1. Factr-I uses the fewest tokens overall but costs a little more than Hermes. Its batched calls produce more output tokens, mostly on OOLONG, and output tokens cost more than input tokens.
  2. Prime rewrote itself in Rust on October 9, 2026. Our benchmarks were ready before that, so we ran the earlier Prime first, then ran the Rust version once it was released. Read Prime’s post.

Releasing soon. Stay tuned.

Join the waitlist

FAQ