From pages to possibilities.
Extract readable Markdown from PDFs and images. Entirely on your machine.
Your document, page by page
Drop a PDF or images above to preview and extract their contents.
Fast mode uses deterministic selection. Speculation is confirmed after generation. Output must fit the remaining context. Changing LM head precision or DFlash2 settings reloads the model on the next request and may repeat warmup. That request can take much longer to show its first token, especially when new compilation is needed. Later requests reuse the prepared engine. INT8/INT4 require single-GPU CUDA, FP16/BF16 and an unquantized, bias-free head. Quantization can change output.
Make yourself understood.
Thoughtful translation. Entirely on your machine.
Bring your words with you.
Import text, subtitles, JSON, DOCX or PDF. Document imports extract text locally; downloads are plain text, without the original document layout.
TXT · MD · SRT · VTT · JSON · DOCX · PDF / up to 8 files, 80,000 characters eachWords worth keeping.
Recent work and your phrasebook, stored only in this browser.
A little knowledge goes a long way.
Ask your documents. Follow the evidence.
Every token, in perspective.
Live hardware readings and measured requests from this workspace.
Decode performance
Recent workspace requests · tok/sRecent requests
Exact generated IDs, including EOS| Model / status | Input | Output | First token | Decode | Speculation |
|---|---|---|---|---|---|
| No requests yet. Start a conversation to see real measurements. | |||||
Decode speed excludes the first token, prefill, warmup, and final output processing. Hardware readings are sampled separately every two seconds. API requests outside this workspace are not listed.
Put your setup to the test.
Compare decoding methods across code, English and translation.
One model. Several ways to run it.
Uses the model selected in Chat, fixed text-only prompts and thinking disabled. Each workload/profile pair gets one discarded warmup, then repeated measurements. A full comparison can take several minutes; switching profiles may reload and compile.
Native heads preserve the model dtype (BF16 for Qwen). DFlash2 precision describes its internal projections; vocabulary heads are labelled separately. The checkpoint controls draft count. Results remain here until page reload. The final profile stays loaded; your next Chat request restores its selected settings.
The first warmup can take several minutes.
Run a benchmark to measure your model.
| Model / workload | Profile | Output cap | Decode tok/s | First token | Acceptance | Runs / status |
|---|---|---|---|---|---|---|
| Choose workloads and profiles, then start the comparison. | ||||||
This measures post-first-token decode speed, not whole-request throughput or quality. Reload, compilation and warmup are excluded from decode speed. First-token latency is from warmed measurements. Early stops or unexpected decoding modes are retained in the report but prevent a complete fixed-length median. Quantized heads may change output and acceptance. Compare the same workload, model, context, output cap and stopping policy. No speedup is inferred across different texts.
One engine. Many possibilities.
Use your local Eole server from other applications.
OpenAI-compatible connection
Point your client at this address and select a configured model.
Chat Completions · Responses API · Anthropic Messages
A quick example
About this workspace
Chats live in this browser's local storage. Clear browser data to remove them. Knowledge collections are temporary server-memory data, isolated by tab session. The server also retains the last 50 workspace request metrics in memory.