Tether · 2026-05-07 · notable
Tether QVAC MedPsy — On-Device Medical LLM Beats Google MedGemma 27B at Less Than One-Sixteenth the Size
Tether's AI group ships open-weights MedPsy-1.7B and MedPsy-4B medical reasoning models that run offline on phones in 1.2–2.6 GB and beat MedGemma-27B on HealthBench Hard. Apache 2.0, GGUF on day one.
A 4B medical model that fits in 2.6 GB and beats Google's 27B MedGemma on hard clinical reasoning — Apache 2.0, runs entirely offline.
Key specs
| Med psy 4 b health bench hard | 58.0 |
|---|---|
| Med gemma 27 b health bench hard | <58 |
| Med psy 1.7 b avg (7 closed ended) | 62.62 |
| Med psy 4 b avg (7 closed ended) | 70.54 |
| Q4 k m size (1.7 b) | 1.2 GB |
| Q4 k m size (4 b) | 2.6 GB |
| Training compute | ~30,000 H100-hours |
What is it?
QVAC MedPsy is a pair of small medical language models (1.7B and 4B) released by QVAC, Tether's AI research group. Both are text-only, open weights under Apache 2.0, and shipped with GGUF builds tuned for llama.cpp on phones, laptops, and edge devices. The 4B variant scores 70.54 across seven closed-ended medical benchmarks (MedQA-USMLE, MedMCQA, MMLU Health, MedXpertQA, PubMedQA, AfriMedQA, MMLU-Pro Health) and posts 58.0 on HealthBench Hard, exceeding MedGemma-27B-text-it.
How does it work?
The team starts from Qwen3 backbones in thinking mode, then runs multi-stage SFT on a broad medical corpus followed by a reasoning-focused corpus, with all reasoning traces generated by a 235B teacher (Baichuan-M3-235B). On top of that they apply two-stage RL — easy/moderate AlphaMedQA cases first, then a hard-enriched RL pass. Recommended Q4_K_M quantization shaves the 4B to 2.72 GB with a -0.81 quality delta. Inference uses 3.2× fewer tokens than the Qwen3 backbone (909 vs 2953 on average).
Why does it matter?
Medical inference today either runs in the cloud (HIPAA headache, latency, cost) or on a 27B model your phone can't hold. MedPsy lands at the intersection both groups have wanted: SOTA-on-HealthBench-Hard performance in a footprint that fits on consumer devices, with permissive licensing. For hospitals, NGOs working in low-bandwidth regions, and any builder who wants medical reasoning without sending PHI to a third party, this resets the local-medical-LLM frontier.
Who is it for?
clinical app builders, edge-AI teams, low-resource health deployments, llama.cpp users
Try it
huggingface.co/qvac/MedPsy-4B-GGUF