QVAC Psy: The Foundation of Stable Intelligence
QVAC Psy, is Tether's family of state-of-the-art foundational models rooted in the principles of Psychohistory. Designed to provide a stable, objective substrate for a decentralized, composable, and infinitely scalable intelligence, QVAC Psy aims to ensure that intelligence remains a constant, unbreakable utility for all without a central authority.
QVAC VisionPsy
QVAC VisionPsy is the second specialized evolution of our foundational logic, purpose-built for edge deployment. These QVAC PSY vision-language models bring multimodal understanding that used to require the cloud onto the phone in your pocket, at just 460M parameters.
Best-in-Class Vision: Across 17 multimodal benchmarks, our 460M model leads its weight class on 16, with the highest overall score of any ~0.5B vision model tested.
Punching Above Its Weight: On ScienceQA, MM-IFEval and POPE, our 460M model outperforms every 0.75B to 1B model tested, including models 2.3 times its size.
Complete Capability Coverage: It leads all four capability areas measured: document understanding and OCR, visual perception, reasoning and knowledge, and instruction following and reliability, with its widest margins in reasoning and visual perception.
Built for Real-Time Vision
Beyond raw accuracy, we have achieved a dramatic reduction in latency. QVAC VisionPsy-Nano-460M-Flash reads an image using as few as 64 visual tokens where comparable models use up to 1088, while retaining ~99% of the full model's quality.
Real-Time Performance: Flash reaches the first token in 0.3 seconds on an iPhone 15 and 2.6 seconds on a Galaxy S25 Ultra, up to 25 times faster than comparable full-token models on the same device.
Lower Memory Footprint: Peak memory drops by 24-39% on Android and 41-65% on iPhone 15 against full-token baselines, freeing the headroom a phone actually has.
Quantized Models: Alongside the full-precision models, we release GGUF quantized variants built for on-device inference through llama.cpp and the QVAC SDK.
QVAC MedPsy
QVAC MedPsy is the first specialized evolution of our foundational logic, purpose-built for edge deployment. These QVAC PSY medical models deliver reasoning capabilities previously exclusive to models seven times their size, setting a new benchmark for efficient, local intelligence.
Unprecedented Efficiency: Across 7 medical benchmarks, our 1.7B model outperforms Google’s MedGemma 4B by over 11 points, despite being less than half its size. In rigorous testing on HealthBench Hard, the 1.7B model also outperformed the nearly sixteen-times-larger MedGemma 27B.
Sovereign Power: On those same 7 medical benchmarks, our 4B model outperformed the nearly seven-times-larger MedgGemma-27B.
Closing the Parameter Gap: This demonstrates that our smaller models, powered by a superior methodology, can match or outperform larger competing state-of-the-art models, achieving top-tier results on real-world medical and health clinical assessments.
Optimized for the Edge
Beyond raw reasoning, we have achieved massive token efficiency. QVAC MedPsy-4B produces accurate answers using 3.2 times fewer tokens than backbone models. Alongside the full-precision models, we also release a full suite of efficient quantized variants.
Real-Time Performance: Lower token counts translate to lower latency and faster inference on edge devices.
Quantized Models: Our 4‑bit quantization retains accuracy within ~1 point of BF16 while reducing disk footprint by ~69%, enabling practical deployment on resource‑constrained edge devices.
Universal Access: Bringing high-level intelligence to bandwidth-constrained and privacy-sensitive environments.
Your Infrastructure,
Your Sovereignty
Every QVAC Psy model is fully open-source, allowing deployment on your own infrastructure.
Full Open Source: Audit, refine, and adapt the weights to your specific needs. All our models, medical and vision, are released in both full-precision checkpoints and complete GGUF-quantized variants.
Local Execution: Achieve near-native speeds on local GPUs using the QVAC Fabric.
Stability by Design: A sovereign engine of intelligence that ensures your data remains private in any environment.
FAQ
MedPsy
MedPsy is a family of compact, text-only medical and healthcare large language models developed by Tether Data’s AI Research for edge and on-device deployment. The family includes MedPsy-1.7B, MedPsy-4B, and GGUF quantized versions for local inference.
All MedPsy models are available in the Hugging Face collection: MedPsy on Hugging Face. The collection includes the full-precision models and GGUF quantized versions for local deployment.
They are intended for developers and researchers building healthcare applications involving medical text, especially privacy-sensitive or on-device use cases. They are starting points for downstream applications and should be validated, adapted, and monitored before production use.
No. MedPsy models are not a substitute for professional medical judgment, clinical diagnosis, or treatment. They can make mistakes, hallucinate, or produce incomplete advice, so outputs should always be reviewed by qualified healthcare professionals where medical decisions are involved.
MedPsy is designed to deliver strong medical reasoning at much smaller model sizes. MedPsy-4B surpassed MedGemma-27B-text-it on the closed-ended benchmark review conducted while being nearly 7x smaller. MedPsy-1.7B is designed for smartphone-class deployment and outperforms larger baselines on several reported medical and HealthBench evaluations.
Use MedPsy-1.7B when memory, latency, or smartphone deployment is the main constraint. Use MedPsy-4B when you want higher quality while still staying within edge-device scale. For most local deployments, the GGUF Q4_K_M variants are the recommended size/quality trade-off based on our testing.
Yes. The GGUF releases are built for local inference through llama.cpp and the QVAC SDK. This enables fully on-device workflows where sensitive health queries do not need to leave the user’s device.
Not if the model is deployed locally through QVAC SDK or another on-device runtime. The models themselves support local inference, but application developers are responsible for ensuring their full product architecture preserves privacy.
For best quality, use Q8_0. For most users, use Q4_K_M with imatrix calibration: about 2.6 GB for MedPsy-4B and 1.2 GB for MedPsy-1.7B. For MedPsy-4B, IQ3_M is a strong around-2 GB option. For MedPsy-1.7B, 3-bit variants are not recommended for medical use.
No. The models were trained and evaluated in English. Performance in other languages has not been validated.
No. MedPsy is text-only. It cannot interpret X-rays, scans, photos, PDFs as images, or other non-text modalities. It can only process information provided as text.
The models are built on Qwen3 backbones and post-trained through a multi-stage medical pipeline: broad supervised fine-tuning, reasoning-focused supervised fine-tuning, and reinforcement learning on medical QA tasks. Long-form reasoning supervision was generated using Baichuan-M3-235B as the teacher model.
They were evaluated across closed-ended medical benchmarks including MedQA-USMLE, MedMCQA, MMLU Health, MMLU-Pro Health, PubMedQA, AfriMedQA, and MedXpertQA, plus HealthBench and HealthBench Hard for realistic open-ended health scenarios.
No. They should never be used in emergency or life-threatening situations. MedPsy models are not a substitute for professional medical judgment, clinical diagnosis, or treatment. They can make mistakes, hallucinate, or produce incomplete advice, so outputs should always be reviewed by qualified healthcare professionals where medical decisions are involved.
The models may hallucinate, miss rare or complex conditions, reflect biases in training data, lack up-to-date medical knowledge, and perform unpredictably. Quantized versions can introduce additional quality degradation, especially at very low bit counts.
The model cards list the models under the Apache 2.0 license. Developers should still review the model cards and applicable terms carefully.
VisionPsy
VisionPsy-Nano is a family of compact vision-language models developed by Tether Data's AI Research for edge and on-device deployment. The family includes VisionPsy-Nano-460M tuned for quality, VisionPsy-Nano-460M-Flash tuned for latency, and GGUF quantized versions for local inference.
Visual question answering, document, chart and diagram understanding, reading text out of a scene, and light visual instruction following. It is built for one image per query.
Use VisionPsy-Nano-460M when accuracy is the priority. Use Flash when time-to-first-token, battery or memory are the binding constraints; it uses as few as 64 visual tokens instead of 1088 and retains ~99% of the full model's quality.
It is designed to deliver strong multimodal understanding at a size that ships. It leads its ~0.5B weight class on 16 of 17 benchmarks, and on ScienceQA, MM-IFEval and POPE it outperforms every 0.75B to 1B model tested, including models 2.3 times its size.
Yes. The GGUF releases are built for local inference through llama.cpp and the QVAC SDK, so images never need to leave the device. Full-precision weights load through Transformers, and vLLM is supported for server-side throughput.
Not if the model is deployed locally through the QVAC SDK or another on-device runtime. The models support local inference, but application developers are responsible for ensuring their full product architecture preserves privacy.
No. It is trained and optimized for one image per query. Multi-image prompts are outside its intended use.
It is primarily English. Other languages are not officially supported yet.
As a compact model it may occasionally hallucinate or miscount, and it is strongest on focused tasks rather than very dense documents or long multi-step maths. It is not intended for safety-critical or high-stakes automated decisions. Because it is small, fine-tuning on your specific domain is the fastest way to raise quality.
Every model was scored in a single VLMEvalKit harness using each benchmark's official metric, with an LLM judge for free-form answers. The eval configs and judge setup are published so the numbers can be reproduced.
The model cards list the models under the Apache 2.0 license. Developers should still review the model cards and applicable terms carefully.
