Specification Sheet
OmniBioFex 1.0 47B is a routed combination of two model heads served together on Groq's LPU stack. A 27B vision head reads the image and extracts structured findings. A 20B text head writes the impression, differential, and management plan. Combined, they total 47B parameters — but they are never stacked in memory, only chained in sequence.
At a Glance
Benchmarks
Benchmark figures are the published numbers for the underlying vision head (qwen/qwen3.8-27b). They describe the model, not the combined pipeline.
Full Specification
| Attribute | Specification |
|---|---|
| Brand | OmniBioFex 1.0 47B |
| Vision head | qwen/qwen3.8-27b · 27B params · bf16 |
| Text head | openai/gpt-oss-20b · 20B params · bf16 |
| Combined parameter count | 47B (sequential routing) |
| Architecture (vision) | Hybrid Gated DeltaNet + Gated Attention |
| Input modalities | Text, images (max 3 per request, 2048 tokens/image) |
| Output modality | Text only (JSON-native) |
| Context length | 131,072 tokens |
| Max output | 16,384 tokens per request |
| Reasoning modes | Thinking + Instruct (switchable per request) |
| Tool use | Function calling + tool orchestration |
| Multilingual | English · Hindi · Tamil · Telugu |
| Inference provider | Groq LPU |
| Streaming | Supported (chat + report assembly) |
| Response format | Strict JSON schema enforcement |
| Scan modes | Standard · Multi-model · Find abnormalities · Compare |
| Templates | General · Chest · Neuro · Orthopedic · Cardiac · Patient-Friendly |
| Spatial grounding | 3×3 region grid per finding |
| Localization verification | Optional 2-pass (adds ~5–8s) |
| Team seats | Up to 10 (Enterprise & Scale tiers) |
| Team roles | Admin · Member · Technician · Reviewer · Billing |
| FHIR output | HL7 FHIR DiagnosticReport webhook (Enterprise) |
| Export formats | PDF · Markdown · JSON |
| Privacy features | Medical Privacy Mode (15-min idle signout, redacted previews) |
Routing Model
A single report run touches both heads, but never at the same time. This keeps memory, latency, and cost predictable.
VISION HEAD
The image is tokenized and sent to the vision head. Output: 6–12 findings with confidence + region, plus a fracture-hunt audit.
TEXT HEAD — IMPRESSION
Findings are passed to the text head. Output: a one-sentence impression and a short advice line.
TEXT HEAD — DIFFERENTIAL
Findings + impression are passed again. Output: a ranked differential and 3–6 management steps.
ASSEMBLY
All three outputs are composed into the final report card and persisted to your private history.
Try it on a real scan
Free first scan. See the full pipeline produce a report in 20–40 seconds.