By continuing you accept the terms of use below.
You are on the machine’s own Wi-Fi. Nothing here goes online, so your phone may warn that this network has no internet — stay connected anyway.
Help us test PhysiAssist
Write your own answers to our questions, correct and rate the assistant’s answers, add questions from your own teaching, or share labelled images of artifacts. Your expertise is what tells us whether it can be trusted.
How it works
Retrieval-augmented generation: the model answers from passages it was handed, not from memory alone.
- 01Search Your question is compared by meaning against all 74,613 passages in the library; the 60 closest are kept.
- 02Rank A second model reads each of those 60 next to your question and picks the best 6.
- 03Answer A language model writes the answer from those passages and cites them as [1]–[6]. Tap a number to see which source it came from.
Everything runs on one NVIDIA DGX Spark. No question is sent to a cloud AI service.
The library
What the answers can draw on. It is fixed — the app does not search the internet.
Of the papers, 2,075 had digital text and 157 older ones were scanned and read by OCR, so those can contain recognition errors — their sources are marked OCR. Indexed July–August 2026; anything published later is not in it.
The models
Three open-weight language models write answers. You can choose one, or let a blind test choose.
Qwen3.8-27B
DEFAULTThe default because in our tests it got the most calculation questions right (49 of 55) and almost always cited its sources. The slowest of the three.
- PARAMETERS
- 27 billion, all used for every token (dense)
- DESIGN
- 64 layers: 48 linear-attention (Gated DeltaNet) + 16 full-attention
- PRECISION
- 8-bit floating point (FP8)
- SPEED HERE
- ~17 tokens/s, about 25 s for a typical answer
Gemma 4 26B-A4B
A mixture-of-experts model: only a small part of it works on each token, so it is fast for its size.
- PARAMETERS
- 25.2 billion total, 3.8 billion active per token
- DESIGN
- 30 layers; 8 of 128 experts + 1 shared expert per token
- PRECISION
- 4-bit (NVFP4)
- SPEED HERE
- ~30 tokens/s
gpt-oss-20b
The fastest, and the highest score on our answer-quality test. But 13% of its answers cited no source, and it more often added numbers that are not in the passages — it leans on its own memory, so check its citations with extra care.
- PARAMETERS
- 21 billion total, 3.6 billion active per token
- DESIGN
- Mixture of experts
- PRECISION
- 4-bit (MXFP4) expert weights
- SPEED HERE
- ~48 tokens/s
Search: BGE-large-en-v1.5
- PARAMETERS
- 335 million
- ROLE
- Step 01: turns text into 1,024-number meaning vectors
Ranking: gte-reranker-modernbert-base
- PARAMETERS
- 149 million
- ROLE
- Step 02: scores each passage against your question
Parameters are the numbers a model learned in training; more of them generally means more knowledge, but each one has to be read from memory for every token written. That is why the 27-billion dense model is the slowest here, and why the mixture-of-experts models, which use under 4 billion per token, are quicker. Fewer bits per parameter (8-bit, 4-bit) makes a model smaller and faster, at a small cost in accuracy.
Settings behind every answer
Fixed for everyone, so answers are comparable.
| Passages given to the model | 6 by default (4–10 selectable) |
| Candidates the ranker reads | 60 |
| Temperature (randomness) | 0.2 — low, mostly consistent |
| Longest answer | 2,000 tokens (8,000 with Thinking on) |
| Most the model can read at once | 16,384 tokens |
| Conversation memory | the last 6 messages |
| Calculation questions | detected automatically; extra instructions to show units and end with a FINAL line |
Options in the app
Along the bottom of the chat screen.
Limit the search to the teaching pages or to the research papers.
How many passages the model is given. More means more evidence, but a slower answer.
Which model writes the answer. A model that is not running is greyed out.
The model reasons privately before answering. Much slower; in our tests it was not more accurate on average.
Picks a model and Thinking setting at random and hides which one answered until you rate it. This is how we compare the models fairly — please use it and rate.
Rate the answer. This is the feedback that decides which model becomes the default.
Speak your question instead of typing it. Click the mic, talk, and click it again. The words appear in the box for you to check before sending. Works on the secure (https) link only; browsers block the microphone elsewhere.
Attach an MRI or CT image (a screenshot is fine) and ask about its physics: the contrast, the sequence, an artifact and how to fix it. The model first writes what it sees (open “What the model saw” under the answer), searches the library with that, then answers with sources. For physics and image quality only, not for diagnosis. Qwen3.8 and Gemma 4 look at the image; gpt-oss reads the notes only.
Tap a number to open the source list and jump to that source.
Ask a new question and it searches again. Say “shorter” or “just the equation” and it revises the last answer instead.
Write your own answers, correct and rate the assistant’s answers, add questions, or add labelled MRI/CT images with known artifacts, so we can test and improve it.
Copy one answer or the whole conversation as text; start over with an empty conversation.
Terms of use
Please read before relying on anything here.
Non-commercial research use only
This is an academic research prototype, offered free to invited colleagues for education and research. It is not a product or service, and it may not be used for any commercial purpose.
Not medical advice
Do not use it for diagnosis, treatment or any patient-care decision. Image questions are for physics and image quality only: the assistant does not read clinical images, and anything it says about an image can be wrong. For MRI safety questions, such as implants and devices, follow the manufacturer’s labelling and your MR safety officer.
Answers can be wrong
The answers are machine-generated. Even with citations they can contain errors, misread sources, or invent details. Check the cited source before you rely on, teach or publish anything from it.
Copyright
The source documents belong to their authors and publishers, including Questions and Answers in MRI (mriquestions.com, © Allen D. Elster). They are used here under access control for non-commercial research. Answers link to the originals; please do not copy or redistribute source material from this app.
Privacy
Your conversation is kept in your browser, not on the machine. When you rate an answer, the question, the answer, which model wrote it, any reasons or comment you add, response details such as the time it took, and an anonymous session code (not your name) are saved to improve the system. The server also logs the times and network addresses of requests. On the Contribute page, what you save (answers, reviews, questions) is stored together with the name you enter, for the project team to review. Images you attach are kept in the assistant's memory for up to an hour so you can ask follow-ups, never saved to disk, and then discarded. Remove patient names, IDs and dates before attaching an image. If you use voice input, the recording is turned into text on the same machine that runs the assistant and then discarded; audio is never stored or sent anywhere else. Never enter patient-identifiable information.
No warranty
Provided as-is, may be changed or taken offline at any time, and comes with no guarantee of accuracy or availability.