MODEL PREVIEW · 2026
ZERO · BENCHMARK-TRAINED LANGUAGE MODEL
Benchmark results.
From actual generations.
One checkpoint, evaluated on the original benchmark prompts.
These canonical prompts and reference answers were included in training; the scores measure recall.
LIVE INFERENCE
Try the model.
Send your own prompt and see an actual model response.
b×
ZeroBenchmarks, measured.
Unverified model output. Novel questions can fail; model-generated claims are not statements by this project.
Real model generation. No answer lookup. No tools.
Ready to load on this device
THE READING LIST
View benchmarks.
Browse measured tasks and source references.
Run available prompts or open their original source.
THE NUMBERS, WITH CONTEXTPerformance.
Performance.
Measured.
Canonical prompts were included in training. One greedy generation per prompt.
Scores measure benchmark recall for the selected release.
| Benchmark | Canonical recall | Surface variation |
|---|
Evaluation protocol
- Decoding
- Greedy · one attempt per prompt
- Training overlap
- Canonical prompts and reference answers
- Answers
- Generated from the actual weights
- Code
- Original tests on generated completions
- Tools / retrieval
- None at inference
Surface-variation counts and methods are recorded separately for each release.
THE EVIDENCEEvery question.
Every question.
On the record.
Inspect the generated outputs behind the scores.
Source statements and evaluation details are available below.
| Task | Expected | Generated | Result |
|---|
Inspect the measured record ↗
PRIVACY & USAGE
Privacy & limits.
Model inference happens on your device.
Your conversations stay here.
Prompts are processed locally and are not sent to our server. The transcript lives in page memory and disappears when cleared or closed. Each message is processed individually. There are no analytics, ads, or third-party inference APIs.
No application cookies.
This static deployment sets no application cookies. A local browser record, bmx_usage, stores only daily and minute usage counts for a device-level demo limit. It is not transmitted to us. Counts reset daily; the record remains until overwritten or you clear site storage. There are no advertising or analytics cookies.
A finite amount of enthusiasm.
Generation runs entirely on your device, with a limit of 10 requests per minute and 1,000 per day in this browser. These local limits can be bypassed and are not a server-wide quota. Hosting usage is governed separately by the hosting provider. Local inference uses your device's CPU, memory, and battery.
Operator and privacy contact.
Benchmaxx project · browser-only inference.
Purpose, retention and your choices.
The application sends no prompts, transcripts, usage counters or model outputs to a server, and stores no server-side user records. The hosting provider necessarily processes network addresses and request metadata to deliver the site; its retention policies apply. Clear site storage to remove local usage counts and cached assets. No signup, profiling, advertising or sale of personal data is part of this application.
For project inquiries, visit the project owner's public profile. The operator may configure an additional privacy contact below. See Vercel's privacy policy for the host's practices.
Third-party software and dataset notices ↗
Model use.
Outputs may be incorrect. See the research disclosure below for the benchmark protocol and its limitations.