Velocity: Proof of Linear-Scaling Long Context Without Model Retraining
Key Takeaways
- ▸MTA Adapt achieves 2.3x faster inference on existing models by selectively attending to relevant tokens, eliminating the need for model retraining
- ▸Velocity runs as a single .exe on consumer GPUs with no Python/PyTorch/cloud dependencies, making local long-context inference more accessible
- ▸The proof build includes transparent benchmarking tools built into the app so users can verify performance claims on their own hardware
Summary
Veloresearch has unveiled Velocity, a local AI execution stack that enables existing language models to handle longer context windows with linear-scaling efficiency—without requiring any model retraining. The system uses MTA (Motify Transit Architecture), a native execution framework that compiles existing models into sealed .mfy artifacts optimized for local CUDA execution on consumer GPUs.
The core innovation is MTA Adapt, an active execution path that selectively scores and attends to relevant tokens in the context cache, maintaining quality while bounding computational cost. In benchmarks, MTA Adapt achieves 2.3x faster decoding at 3.5k context compared to the exact (full-window) baseline, while maintaining bit-identical quality at short context lengths. At 32k context, per-token attention costs drop 30x+ below the exact computation path.
Veloresearch has released a proof build for Windows as a single executable with no Python, PyTorch, server, or cloud dependencies. The package includes a self-contained model artifact (Qwen 3.5 4B), built-in benchmark tools for local verification, and an Exact mode for auditing baseline behavior—emphasizing reproducibility and transparency over unverified claims.
- Quality remains bit-identical at short context and computational cost scales linearly with context length—a narrow but rigorously measured technical claim
Editorial Opinion
Velocity represents a pragmatic approach to long-context inference on consumer hardware—rather than overpromising a universal solution, the team has grounded their claims in verifiable benchmarks built directly into the product. The emphasis on transparency (Exact mode for auditing, /bench commands for local verification) stands out in a space often dominated by closed implementations. If the linear-scaling properties hold across diverse hardware ecosystems, this could meaningfully lower the barrier to running sophisticated local models, though the 4B model size remains a fundamental limitation for complex reasoning tasks.
