Velocity a proof of linear-scaling long context for existing LLMs, no retraining Velocity, a startup building the Motify execution stack for local AI models, released a Windows proof build of its MTA Adapt technology that claims linear-scaling long context for existing large language models without retraining. In back-to-back tests on an NVIDIA RTX 3060 Laptop GPU, MTA Adapt achieved 58.6 tokens per second at 3.5k context versus 25.6 tok/s for the exact path, a 2.3x speedup, while maintaining near-identical perplexity (9.101 vs 9.028 at 4096 positions) and bit-identical results at 1024 positions. The build, which runs as a single .exe with no Python or cloud dependencies, includes a benchmark suite that lets users verify the claims on their own hardware. Velocity is not another AI chat app. Velocity builds Motify — a native execution stack for local AI models, based on sealed .mfy artifacts and MTA Motify Transit Architecture . The chat is only the interface; the execution stack underneath is the product. This is a proof build : everything it claims, it can measure on your machine. existing model → MTA compiler → sealed .mfy artifact → Velocity runtime → CUDA execution path → MTA Exact / MTA Adapt → local AI No Python. No PyTorch. No server. No cloud. One .exe . Windows proof build | | Model artifact veloresearch/qwen3.5-4b-adapt-b32 on Hugging Face https://huggingface.co/veloresearch/qwen3.5-4b-adapt-b32 The installer downloads the .mfy artifact ~2.95 GB during setup, SHA-256 verified. If skipped, velocity.exe downloads and verifies it on first launch, with resume support for interrupted downloads. No account or token required. Windows SmartScreenwill warn about an unknown publisher — the build is not code-signed yet. ClickMore info → Run anyway. MTA Adapt matches the exact path — and beats it as context grows. Measured back-to-back on the same machine, same model, same text WikiText-2 raw test sample : Quality ppl @1024 positions : EXACT 11.312 = ADAPT 11.312 0.0% — bit-identical ppl @4096 positions : EXACT 9.028 vs ADAPT 9.101 +0.8% Speed decode @3.5k context : ADAPT 58.6 tok/s vs EXACT 25.6 tok/s 2.3x faster decode, short context: ~52–55 tok/s both paths per-token attention @32k context: 30x+ below the full-window path Below ~2k context Adapt attends the entire window — it is literally the exact computation . As context grows, it selects the active keys and the cost stays bounded while Exact's grows linearly. That crossover is the product. Not a universal claim — a narrow, verifiable one. The benchmark suite ships inside the app: type /bench and it renders these exact charts from your hardware. If Adapt loses on your GPU, the chart will show it. GPU NVIDIA RTX 3060 Laptop GPU, 6 GB VRAM Backend CUDA, GPU-resident Q4 path Artifact qwen3.5-4b-adapt-b32.mfy frozen Qwen-family 4B, 4-bit OS Windows 11 x64 VRAM ~3 GB total at the default context budget A mid-range consumer laptop GPU, deliberately. If it runs here, it runs on ordinary hardware. Your numbers will differ with GPU, drivers and thermals — don't quote ours; measure yours. | Path | Status | Role | |---|---|---| MTA Exact | working | full-window reference path — baseline, parity, auditability | MTA Adapt | working | the product path — active execution over frozen weights, no retraining | MTA Native | future | models designed directly for Velocity's execution stack | Exact proves trust. Adapt ships existing models. Native breaks the ceiling. The verification path. Use it to answer one question: does this .mfy artifact preserve expected baseline behavior? Every optimized mode should be judged against a reference the user can run — Exact is that reference, one command away /mode exact .The product path. It runs an existing, frozen model through active execution: an embedded selector scores the cached context cheaply, exact-rescores the best candidates, and attention runs over that active set — sink, recent window, and selected keys. The active budget adapts to the context, so quality holds while the cost stays bounded. No retraining, no calibration; the selector ships inside the artifact. Verify it yourself, one command each: /mode exact → /bench ppl /mode adapt → /bench ppl A .mfy file is a sealed Motify model artifact: model payload, tokenizer metadata, runtime configuration and the embedded Adapt selector in one portable file. Hugging Face stores it as a regular binary; Velocity is the runtime that knows how to open and execute it. qwen3.5-4b-adapt-b32.mfy ~2.95 GB, self-contained — nothing else to install Velocity shows its execution surface instead of hiding it. /map on renders, per layer, live during generation: which path each layer runs, the measured active-KV ratio , context usage, and prefill/decode speed. | Term | Meaning in the map | |---|---| GQA | Grouped Query Attention — the attention path; bars sized by measured active-KV % | KV | Key/Value cache — attention memory; grows with context on the Exact path | SSM | State Space Model — layers with a compact working state instead of a growing window | O 1 state | a bounded working state that does not grow with conversation length | FFN | Feed-Forward Network — the dense compute block in each layer | We believe in runnable proof, not screenshots. /bench full measures your machine, renders shareable PNG charts each stamped with your device name and writes the raw numbers to summary.txt . Nothing is baked in, Exact is always plotted next to Adapt, and every step prints its wall time. Reference results from the tested configuration: | Chart | | |---|---| | Context speedup /Veloresearch/velocity-mta-proof/blob/main/benchmarks/02 speedup.png Kernel bandwidth /Veloresearch/velocity-mta-proof/blob/main/benchmarks/03 kernel bandwidth.png measured read ceiling of the GPU Decode throughput /Veloresearch/velocity-mta-proof/blob/main/benchmarks/04 decode throughput.png Perplexity /Veloresearch/velocity-mta-proof/blob/main/benchmarks/05 perplexity.png The perplexity corpus is the opening ~120 KB of the WikiText-2 raw test split the same family as llama.cpp's wiki.test.raw , embedded so the number is reproducible offline. It is a sample of the split — treat it as an Exact-vs-Adapt comparison on standard text, not a paper-comparable full-WikiText-2 score. /bench ppl