{"slug": "velocity-a-proof-of-linear-scaling-long-context-for-existing-llms-no-retraining", "title": "Velocity a proof of linear-scaling long context for existing LLMs, no retraining", "summary": "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.", "body_md": "Velocity is not another AI chat app. Velocity builds **Motify** — a native execution stack for\nlocal AI models, based on sealed `.mfy`\n\nartifacts and **MTA (Motify Transit Architecture)**.\nThe chat is only the interface; the execution stack underneath is the product.\n\nThis is a **proof build**: everything it claims, it can measure on your machine.\n\n```\nexisting model → MTA compiler → sealed .mfy artifact → Velocity runtime\n             → CUDA execution path → MTA Exact / MTA Adapt → local AI\n```\n\nNo Python. No PyTorch. No server. No cloud. One `.exe`\n\n.\n\nWindows proof build |\n|\n\n**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`\n\nartifact (~2.95 GB) during setup, SHA-256 verified. If\nskipped, `velocity.exe`\n\ndownloads and verifies it on first launch, with resume support for\ninterrupted downloads. No account or token required.\n\nWindows SmartScreenwill warn about an unknown publisher — the build is not code-signed yet. ClickMore info → Run anyway.\n\n**MTA Adapt matches the exact path — and beats it as context grows.** Measured back-to-back on\nthe same machine, same model, same text (WikiText-2 raw test sample):\n\n```\nQuality   ppl @1024 positions : EXACT 11.312 = ADAPT 11.312   (0.0% — bit-identical)\n          ppl @4096 positions : EXACT 9.028 vs ADAPT 9.101    (+0.8%)\n\nSpeed     decode @3.5k context : ADAPT 58.6 tok/s vs EXACT 25.6 tok/s   (2.3x faster)\n          decode, short context: ~52–55 tok/s both paths\n          per-token attention @32k context: 30x+ below the full-window path\n```\n\nBelow ~2k context Adapt attends the entire window — it is **literally the exact computation**.\nAs context grows, it selects the active keys and the cost stays bounded while Exact's grows\nlinearly. That crossover is the product.\n\nNot a universal claim — a narrow, verifiable one. The benchmark suite ships inside the app:\ntype `/bench`\n\nand it renders these exact charts from **your** hardware. If Adapt loses on your\nGPU, the chart will show it.\n\n```\nGPU       NVIDIA RTX 3060 Laptop GPU, 6 GB VRAM\nBackend   CUDA, GPU-resident Q4 path\nArtifact  qwen3.5-4b-adapt-b32.mfy (frozen Qwen-family 4B, 4-bit)\nOS        Windows 11 x64\nVRAM      ~3 GB total at the default context budget\n```\n\nA 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.\n\n| Path | Status | Role |\n|---|---|---|\nMTA Exact |\nworking | full-window reference path — baseline, parity, auditability |\nMTA Adapt |\nworking | the product path — active execution over frozen weights, no retraining |\nMTA Native |\nfuture | models designed directly for Velocity's execution stack |\n\nExact proves trust. Adapt ships existing models. Native breaks the ceiling.\n\nThe verification path. Use it to answer one question: *does this .mfy artifact preserve\nexpected baseline behavior?* Every optimized mode should be judged against a reference the user\ncan run — Exact is that reference, one command away (\n\n`/mode exact`\n\n).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.\n\nVerify it yourself, one command each:\n\n```\n/mode exact  →  /bench ppl\n/mode adapt  →  /bench ppl\n```\n\nA `.mfy`\n\nfile is a sealed Motify model artifact: model payload, tokenizer metadata, runtime\nconfiguration and the embedded Adapt selector in one portable file. Hugging Face stores it as a\nregular binary; Velocity is the runtime that knows how to open and execute it.\n\n```\nqwen3.5-4b-adapt-b32.mfy   (~2.95 GB, self-contained — nothing else to install)\n```\n\nVelocity shows its execution surface instead of hiding it. `/map on`\n\nrenders, per layer, live\nduring generation: which path each layer runs, the measured **active-KV ratio**, context usage,\nand prefill/decode speed.\n\n| Term | Meaning in the map |\n|---|---|\nGQA |\nGrouped Query Attention — the attention path; bars sized by measured active-KV % |\nKV |\nKey/Value cache — attention memory; grows with context on the Exact path |\nSSM |\nState Space Model — layers with a compact working state instead of a growing window |\nO(1) state |\na bounded working state that does not grow with conversation length |\nFFN |\nFeed-Forward Network — the dense compute block in each layer |\n\nWe believe in runnable proof, not screenshots. `/bench full`\n\nmeasures **your** machine, renders\nshareable PNG charts (each stamped with your device name) and writes the raw numbers to\n`summary.txt`\n\n. Nothing is baked in, Exact is always plotted next to Adapt, and every step prints\nits wall time.\n\nReference results from the tested configuration:\n\n| Chart | |\n|---|---|\n|\n\n[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\nfamily as llama.cpp's `wiki.test.raw`\n\n), embedded so the number is reproducible offline. It is a\nsample of the split — treat it as an Exact-vs-Adapt comparison on standard text, not a\npaper-comparable full-WikiText-2 score. `/bench ppl <file>`\n\nscores any text you choose.\n\n**1. Install** — run [ VeloSetup.exe](https://github.com/Veloresearch/velocity-mta-proof/releases/latest).\nPer-user, no administrator prompt. The model downloads during setup (or on first launch).\n\n**2. Launch** — start **Velocity** from the Start menu (Windows Terminal recommended).\n\n**3. Verify the proof:**\n\n```\n/mode exact      run the reference path\n/mode adapt      run active execution\n/bench           measure both on your machine (charts + summary.txt)\n/map on          watch the per-layer execution map live\n/mode adapt | exact          switch MTA execution path\n/backend auto | cuda | cpu   select the compute backend\n/ctx | /ctx <tokens>         context budget — VRAM estimate per size, applies live\n/think on | off              let the model reason before answering\n/map on | off                per-layer MTA execution map\n/bench                       benchmark menu: quick | ppl | speed | full (PNG charts)\n/bench ppl <file>            score perplexity on your own text file\n/stats                       last-turn speed and context stats\n/new  /copy  /save <file>    conversation & code-block helpers\n/settings  /help  /exit\n```\n\n`Ctrl+C`\n\nstops the current generation without closing the app.\n\n```\nvelocity.exe --model <path.mfy>      use a specific artifact\nvelocity.exe --backend auto|cuda|cpu\nvelocity.exe --exact                 start in MTA Exact\nvelocity.exe --max-ctx N             context budget (default 8192; raise with VRAM headroom)\nvelocity.exe --max-new N             max answer tokens (default 8192)\nvelocity.exe --think                 enable model reasoning by default\nvelocity.exe --plain                 no fullscreen UI / colors\nvelocity.exe --prompt \"...\"          one-shot answer, then exit\n```\n\n**Context budget defaults to 8192 tokens**— a deliberate VRAM choice for 6 GB cards, not an architecture ceiling. Raise it live with`/ctx 16384`\n\n(shows the VRAM estimate first).**Windows x64 + NVIDIA CUDA**(GTX 16xx / RTX 20xx or newer) is the tested performance path. A native CPU (AVX2) fallback runs the same artifact, slower.**Greedy decoding, model as-is**— no sampling tricks, no anti-repeat rewriting on top.** One public artifact so far**(Qwen-family 4B). A Gemma-family artifact is in internal validation. MTA Native is the roadmap headline.\n\n```\nMTA Exact        working\nMTA Adapt        working — bit-identical to Exact below ~2k ctx, active execution above\n.mfy artifact    working\nCUDA backend     working — preferred path\nCPU x86 backend  working — compatibility fallback\nHF auto-download working — setup-time and first-launch, SHA-256 verified, resume\nLocal chat       working\nBenchmark suite  working — /bench, reproducible locally\nExecution map    working — /map\nGPU-vs-CPU check working — /verify (exact, adapt and FFN paths)\n```\n\nVelocity does not compete with Qwen, Gemma or Llama. It builds the execution layer underneath\nthem. If existing model families can be compiled into `.mfy`\n\nartifacts, verified through MTA\nExact, and executed through MTA Adapt without retraining — the value is not in any one model.\nIt is in the artifact standard and the runtime.\n\nAnd the proof runs on a 6 GB consumer laptop GPU.\n\n- not another chat UI, prompt wrapper, or hosted API skin\n- not a fine-tuning product\n- not a cloud demo\n- not a claim without a runnable local proof\n\nInference is fully local. The only network access is the one-time model download from Hugging Face. Conversations stay on your machine.\n\nThis repository contains a public proof build and packaged binaries for Velocity / Motify.\nVelocity, Motify, MTA, the `.mfy`\n\nartifact format, and the runtime/compiler technology are\nproprietary Velocity technologies unless explicitly stated otherwise. Model artifacts follow the\nlicense of their upstream base models. This repository does not grant permission to copy, modify,\nredistribute, or reverse engineer Velocity proprietary technology.\n\n© 2026 Velocity / Velo Research. All rights reserved.\n\n- Website:\n[veloresearch.com](https://veloresearch.com/) - Contact:\n[contact@veloresearch.com](mailto:contact@veloresearch.com) - Artifact:\n[veloresearch/qwen3.5-4b-adapt-b32](https://huggingface.co/veloresearch/qwen3.5-4b-adapt-b32)\n\n```\nDownload it. Run it. Verify it. Break it.\n```\n\n", "url": "https://wpnews.pro/news/velocity-a-proof-of-linear-scaling-long-context-for-existing-llms-no-retraining", "canonical_source": "https://github.com/Veloresearch/velocity-mta-proof", "published_at": "2026-08-02 12:17:09+00:00", "updated_at": "2026-08-02 12:52:56.565838+00:00", "lang": "en", "topics": ["artificial-intelligence", "machine-learning", "large-language-models", "ai-infrastructure", "ai-research"], "entities": ["Velocity", "Motify", "MTA Adapt", "MTA Exact", "NVIDIA RTX 3060 Laptop GPU", "Qwen3.5-4b-adapt-b32", "Hugging Face", "Windows 11"], "alternates": {"html": "https://wpnews.pro/news/velocity-a-proof-of-linear-scaling-long-context-for-existing-llms-no-retraining", "markdown": "https://wpnews.pro/news/velocity-a-proof-of-linear-scaling-long-context-for-existing-llms-no-retraining.md", "text": "https://wpnews.pro/news/velocity-a-proof-of-linear-scaling-long-context-for-existing-llms-no-retraining.txt", "jsonld": "https://wpnews.pro/news/velocity-a-proof-of-linear-scaling-long-context-for-existing-llms-no-retraining.jsonld"}}