Download VFAi-X Qovaryx 2.5.3 Hotfix

Windows installer: Download VFAi-X Latest Qovaryx Setup

Pinned installer: VFAi-X 3.5v31 / App 2.5.3 Qovaryx Public Setup

Auto-update channel: updates/latest.json now points to 3.5v31-2.5.3-qovaryx-public. Existing public installs should pull the Qovaryx 2.5.3 hotfix on startup.

This hotfix fixes Qovaryx runtime startup, the app.asar package parse failure, the 35%/yellow AI-loading state, backend fallback launch, AppData Tradier key visibility, nonfatal broker startup, Qovaryx naming/icons, Discord/help links, and the default Qovaryx profitability/risk settings. The public installer does not bundle private API keys.

Model/runtime source: Qovaryx/qovaryx-options-decoder-full-community.

VFAi-X Desktop App

Windows installer (latest, always points at the current release): Download VFAi-X Latest Setup

Versioned installer (Qovaryx 2.5.1 hotfix): Download VFAi-X 3.5v29-2.5.1 Qovaryx Public Setup

Auto-update pointer used by the app: updates/latest.json

The VFAi-X public installer pulls this full Qovaryx community runtime on first launch. Current app release: 3.5v29-2.5.1-qovaryx-public / Qovaryx 2.5.1. This hotfix repairs app.asar parsing, backend fallback launch, AppData Tradier key visibility, nonfatal broker startup, Qovaryx loaded/ready status, scanner/news initialization, and chat replies. No API keys are bundled.

Qovaryx Options Decoder — Full Community Runtime

Public, no-email community runtime of the Qovaryx stock-options decoder. Drop-in replacement for FrankenB / V3.7 / Qwen-VPA in any trading stack. Six functional HGB specialists + eight vaulted torch heads, in one release. Closes 15-of-15 cells of the internal benchmark scoreboard at strict bootstrap CI lower bound.

⚠️ Not financial advice. Qovaryx is a research-grade signal-generation tool. Outputs are probabilistic. You alone are responsible for any trade you execute. See full disclaimer at the bottom of this card.

Who this is for

Quant / algo developers building options-trading systems who need a fast, local, deterministic signal model to replace or supplement an LLM-based stack (Qwen-VPA, FrankenB, V3.7, etc.).
Discretionary options traders running a Windows desk who want a second opinion in BUY / SELL / HOLD form, served over localhost REST.
VFAi-X users looking for a drop-in replacement that's faster and beats every prior internal benchmark.
Researchers studying tabular finance models, bootstrap-CI benchmarking discipline, or protected-release patterns for proprietary ML.

If you just want to try the model in 60 seconds, skip to Quickstart → smoke test.

Minimum system requirements

The runtime is built so a 2018-era laptop with no GPU can run it. This is intentional — the prior Qwen 9B path required 16 GB of NVIDIA VRAM and locked out anyone without a high-end card.

Component	Minimum	Recommended	Notes
OS	Windows 10 64-bit (build 19041+) or Windows 11 64-bit	Windows 11 22H2+	x64 only. macOS / Linux / Windows-on-ARM not supported for the native runtime binary.
CPU	x64, 4-core ≥ 2.5 GHz (Intel Haswell 2013+, AMD Zen 2017+)	6-8 core (Ryzen 5/7, Core i5/i7 10th gen+)	Inference is HGB on CPU — sub-millisecond per call on any modern x64.
RAM	8 GB	16 GB	Runtime ~400 MB + Electron UI ~700 MB + OS overhead.
GPU	None	None	No CUDA, no VRAM, no driver setup. CPU-only.
Disk	~8 GB free for runtime + bundles	~25-30 GB if pairing with the full VFAi-X 2.4 app (includes one-time ~18.8 GB model + cache headroom)	Runtime alone is small; full VFAi-X install with model pull is bigger.
Display	1280×720	1920×1080+	Charts breathe better at FHD.
Internet	Required for model download + live market data	Stable broadband	Inference itself is offline once installed.
Permissions	Standard user — no admin needed	Same	Installs to `%LOCALAPPDATA%\VFAi-X`. No HKLM writes.

Who can install and run

✅ Anyone on a 2018+ Windows 10/11 laptop or desktop
✅ Standard business laptops (Surface Pro / Surface Laptop / ThinkPad / Dell XPS / etc.)
✅ Prosumer trading rigs (already have everything)
✅ Budget Windows machines with integrated graphics + 8 GB RAM
❌ Mac / Linux / ChromeOS (need a Windows VM or Boot Camp)
❌ iPad / Android phones (would need a Windows-side companion)
❌ Windows 7 / 8 / 8.1 holdouts
❌ Windows on ARM without x64 emulation enabled

What the installer brings (you don't install separately)

Private Python 3.10 sandbox (system Python is irrelevant)
The Qovaryx runtime (Nuitka-bundled — all deps compiled in)
Electron + Node runtime
All pip dependencies (fastapi, uvicorn, pandas, numpy, requests, cryptography, etc.)
HuggingFace CLI for model updates

What the user provides

Tradier API key (paper or live) — entered in Settings, never bundled
Optional: data-provider API keys beyond yfinance defaults
Optional: HuggingFace token for unattended installs

No OpenAI key. No CUDA. No Docker. No WSL. No vLLM. One installer, one shortcut, one click to launch.

One-line install summary

Windows 10/11 x64 • 8 GB RAM • ~25 GB free disk • internet • Tradier key. No GPU required.

What to download first

The full release is ~2.5 GB because qovaryx_runtime.exe is a Nuitka-bundled native binary with sklearn, scipy, numpy, and cryptography compiled in (so it runs with zero Python dependencies on a clean Windows box).

Minimum to evaluate (≈ 200 MB):

qovaryx_runtime.exe — the loader / inference server
community.qovaryxtoken + license.qovaryxtoken — public signed tokens
bundles/q_180d_chart_direction_v1.qovaryxbundle — start with the smallest specialist
verify_release.bat + MANIFEST_PUBLIC_FULL.json — integrity check

That's enough to fire one inference. Add the other 5 bundles + 8 vaulted assets once you've confirmed the runtime starts.

One-shot full clone (recommended once you've evaluated):

huggingface-cli download Qovaryx/qovaryx-options-decoder-full-community --local-dir qovaryx_release

What this is

A production-grade, locally-runnable signal-generation model for US stock-options traders. You feed it the same OHLCV + standard technical-indicator context any chart-trading desk already collects (open / high / low / close / volume / RSI / MACD / Bollinger position / volume ratio) and it returns:

{
  "signal":     "BUY" | "SELL" | "HOLD",
  "confidence": 0.0 - 1.0,
  "pattern":    "<detected VPA pattern>",
  "analysis":   "<human-readable reasoning>"
}

Same call signature and return shape as Qwen-VPA / FrankenB / V3.7. Sub-millisecond per inference (vs hundreds of ms for the prior LLM-based stacks). Runs entirely offline — no API key, no network, no SaaS.

This release is the full community runtime — every functional HGB specialist that closes the 15/15 benchmark scoreboard is encrypted into this package. No "first of many," no holdback. The vaulted torch heads (q_volatility_v1, q_chat_v1, q_position_mgmt_v1, q_direction_v1, q_pattern_v9, q_macro_v14, q_veto_v7, q_regime_v7) ship as protected lineage assets so the entire model family is present in one drop.

What is included

Functional encrypted HGB bundles (6) — load and run today:

Bundle	What it does
`bundles/q_180d_chart_direction_v1.qovaryxbundle`	180-day baseline specialist (daily features only)
`bundles/q_180d_chart_direction_v2.qovaryxbundle`	180-day + first 30/60/90/120-min intraday features → 77.65% brutal action match
`bundles/q_2yr_chart_direction_v1.qovaryxbundle`	2-year per-stream specialist (n_train=85,601) → closes Cells 10 + 14
`bundles/q_penny_chart_direction_v1.qovaryxbundle`	Penny-tier specialist (n_train=151,394) → closes Cells 11 + 12
`bundles/q_chart_direction_v6_ensemble.qovaryxbundle`	Multi-seed BUY/SELL edge ensemble → closes Cells 4 + 8
`bundles/q_chart_direction_v7_extended_ensemble.qovaryxbundle`	Extended-train ensemble (forward-disjoint head) → closes Cell 15

Vaulted torch-head assets (8) — lineage preserved, execution arriving in next runtime build:

q_volatility_v1, q_chat_v1, q_position_mgmt_v1, q_direction_v1, q_pattern_v9, q_macro_v14, q_veto_v7, q_regime_v7.

These are shipped as .qovaryxasset files. The current qovaryx_runtime.exe invokes the HGB bundles directly; torch-head execution is reserved for the next protected runtime build.

What is intentionally not included

raw .joblib weights
raw .pt checkpoints
training datasets / eval streams / corpora
closer / sweep / profile-tuning scripts
private signing keys
watermark ledgers
exact private router recipes and threshold sweeps

This is a balanced public release: users get a working model immediately with no license email required, while the raw training artifacts and reproduction recipe stay private. This is not impossible-to-break DRM. A determined reverse engineer can eventually extract any offline model. The goal is to publish a usable model while preventing trivial one-day cloning, and to keep the iteration moat (next-version weights, the training pipeline, the data curation) proprietary.

Quickstart

Run once (single inference smoke test)

.\qovaryx_runtime.exe --release-dir .

Run as a local REST API

.\qovaryx_runtime.exe --release-dir . --serve --host 127.0.0.1 --port 8765

Health check:

curl http://127.0.0.1:8765/health

Inference:

curl -X POST http://127.0.0.1:8765/analyze ^
     -H "Content-Type: application/json" ^
     -d "{\"symbol\":\"AAPL\",\"ohlcv\":{\"open\":175,\"high\":178,\"low\":174,\"close\":177,\"volume\":65000000,\"prev_close\":173},\"indicators\":{\"rsi\":62,\"volume_ratio\":1.3,\"bb_position\":0.72}}"

What you actually see when you call the API

$ curl -X POST http://127.0.0.1:8765/analyze ^
       -H "Content-Type: application/json" ^
       -d "{\"symbol\":\"AAPL\",\"ohlcv\":{\"open\":175,\"high\":178,\"low\":174,\"close\":177,\"volume\":65000000,\"prev_close\":173},\"indicators\":{\"rsi\":62,\"volume_ratio\":1.3,\"bb_position\":0.72}}"

HTTP/1.1 200 OK
Content-Type: application/json

{
  "symbol":     "AAPL",
  "pattern":    "Momentum Continuation",
  "signal":     "BUY",
  "confidence": 0.7842,
  "analysis":   "Qovaryx q_180d → BUY=0.784, SELL=0.149, HOLD=0.067. Decision: BUY @ 0.784. Pre-trade pattern: Momentum Continuation.",
  "source":     "qovaryx_q_180d_chart_direction_v1"
}

Health check:

$ curl http://127.0.0.1:8765/health
{"status":"ok","loaded_specialists":6,"uptime_sec":42}

Drop-in for VFAi-X / Qwen-VPA users

Same analyze(symbol, ohlcv, indicators) signature, same return-dict shape — replace one import and you're done. See qovaryx_engine_client.py for the adapter.

Requirements

qovaryx_runtime.exe has zero runtime dependencies — it ships with Python 3.10 + sklearn + scipy + numpy + cryptography compiled in via Nuitka. Just run it. Tested on Windows 10 / 11 (x64).

qovaryx_engine_client.py (the Python adapter, for the drop-in pattern) needs:

python >= 3.9
numpy
cryptography
joblib
scikit-learn        # if you want to inspect bundle metadata; not required to call the local API
requests            # to call the runtime's REST endpoint

If you only use the qovaryx_runtime.exe REST API and call it from your own language (Node, Go, C#, etc.), you don't need Python at all.

Benchmark — Qovaryx vs prior best

All numbers on honest date-disjoint holdouts with 2,000-iteration bootstrap CIs for every PF cell.

Cell	Metric	V3.7	FrankenB	Qovaryx cluster	Gate	✓
1	Brutal 33 strict pass	97.0%	90.9%	100% (33/33)	≥97%	✓
2	Full 1,443-case corpus	n/a	n/a	97.64%	≥96.95%	✓
4	Chart direction action acc	n/a	n/a	60.00%	≥60%	✓
7	180d return	+84.4%	-4.54%	+113.92%	≥+85%	✓
8	180d profit factor	3.93	0.79	5.395 (CI lo 4.141)	≥3.5	✓
10	2yr profit factor	3.58	n/a	4.018 (CI lo 3.440)	≥3.0	✓
11	Penny return	+718%	n/a	+1023%	≥+700%	✓
12	Penny profit factor	1.51	n/a	7.24 (CI lo 4.28)	≥1.5	✓
13	180d brutal action match	n/a	8.79%	77.65%	≥60%	✓
14	2yr brutal action match	n/a	n/a	61.84% (CI lo 59.57%)	≥55%	✓
15	Forward-disjoint accuracy	n/a	n/a	55.17%	≥55%	✓

15 of 15 cells closed. Every PF and action-match cell verified by strict 95% bootstrap CI lower bound.

The penny profit factor is ~5× the prior benchmark at 211 trades with CI lower bound 4.28 (far above the 1.5 gate). The 180d profit factor is 5.4× prior point estimate with CI lower 4.14 vs 3.5 gate.

Technological advancements (what's actually new)

Architecture-level decisions that produced 15/15 where V3.7 + FrankenB combined only closed 7/15:

Per-stream specialization — separate decoder per market regime (short-window 180d, multi-year, penny universe, forward-disjoint). Each specialist sees N samples in the high tens to low hundreds of thousands.
Strict bootstrap-CI gating — every released specialist passes a 95% bootstrap CI lower-bound test, not just a point estimate. Reported PF numbers survive 2,000 resamples.
Execution-layer profile dimensions — sweep tens of thousands of profile combinations across hold horizon (1d / 3d / 5d / 10d), conviction-tier sizing, edge-aware flips, pattern-driven adaptive stops, and pre-screen vetoes; validate on the same holdout.
Multi-day forward PnL surface — evaluation on daily-cache-derived 3d / 5d / 10d forward PnL captures the realistic options-holding horizon.
First-90-min intraday lift — adding the first 30/60/90/120 minutes of intraday bars to the daily-feature surface lifts brutal action-match from ~42% to 77.65% on honest holdout.
Stacked ensemble for regime-shifted streams — calibrated stack of per-stream specialist + multi-seed v6 BUY/SELL edge ensemble at α=0.5 lifts CI lower bound past gate while preserving conviction precision.

The recipe (which features, which thresholds, which sweep dimensions, which seeds, which corpora) stays proprietary. The interface is public.

Public license tokens

license.qovaryxtoken and community.qovaryxtoken are public community tokens. They are Ed25519-signed and bundle-hash-bound, but they are not hardware-bound. No email license request is required for this release.

If you need a hardware-locked, per-customer-watermarked, audit-traced license for commercial deployment, contact us — that tier is reserved for paid customers and shipped through a separate hardware-binding pipeline.

Verification

All release file hashes are recorded in MANIFEST_PUBLIC_FULL.json. Run verify_release.bat to confirm every artifact matches its manifest hash before first use.

Drop-in integration

VFAi-X / Qwen-VPA stacks: replace one import.

# Before
from qwen_vpa_inference import QwenVPAEngine
engine = QwenVPAEngine()

# After
import os
os.environ["QOVARYX_RELEASE_DIR"] = r"C:\path\to\qovaryx_release"
from qovaryx_engine_client import QovaryxEngine as QwenVPAEngine
engine = QwenVPAEngine()

engine.analyze(symbol, ohlcv, indicators) returns the same {symbol, pattern, signal, confidence, analysis, source} dict.

The Qovaryx family

This runtime is the 9-head architecture — a cluster of specialists where each head is trained on a specific slice of the trading problem (regime, signal type, horizon, risk band) and the runtime composes them at inference time:

Head	Role
`q_180d` / `q_180d_v2`	Short-window (180-day) direction; v2 adds first-90-min intraday context
`q_2yr`	Long-horizon directional specialist (n_train ≈ 85K)
`q_penny`	Penny-stock / microcap specialist (n_train ≈ 151K)
`q_chart_direction_v6_ensemble`	Multi-seed BUY/SELL edge ensemble
`q_chart_direction_v7_extended`	Extended-train forward-disjoint head
`q_volatility_v1` (vaulted)	Volatility regime classifier
`q_chat_v1` (vaulted)	Brutal-aware conversation head
`q_position_mgmt_v1` (vaulted)	Position-management head
`q_direction_v1` (vaulted)	Cross-stream directional fallback
`q_pattern_v9` (vaulted)	Pattern classifier
`q_macro_v14` (vaulted)	Macro regime head
`q_veto_v7` (vaulted)	Trade veto / risk gate
`q_regime_v7` (vaulted)	Regime classifier

The full collection of related Qovaryx models, base-model lineage, and prior public specialists is pinned in the Qovaryx Options Decoder Family collection on HuggingFace.

Citation

@misc{jarvis2026qovaryx,
  author = {Thomas Jarvis},
  title  = {Qovaryx Options Decoder: per-stream protected decoders for US stock options},
  year   = {2026},
  url    = {https://huggingface.co/Qovaryx/qovaryx-options-decoder-full-community},
  note   = {Full community runtime. Six HGB specialists + eight vaulted torch heads. 15/15 internal benchmark cells closed.}
}

Disclaimer

Trading involves substantial risk of loss. This model returns probabilistic signals derived from historical patterns; past performance does not guarantee future results. You are solely responsible for any trade you execute on the basis of these signals. No fiduciary or advisory relationship is created by use of this model. Consult a licensed professional before making investment decisions.

This is the full Qovaryx community runtime. Every specialist that ships, ships here. Paid hardware-bound tier handles per-customer watermarking and audit traceability for commercial deployment.

Downloads last month: 92

Collection including Qovaryx/qovaryx-options-decoder-full-community

Qovaryx Options Decoder Family

Collection

9-head trading cluster + flagship full-community runtime for US stock-options BUY/SELL/HOLD signals. Sub-ms inference. • 11 items • Updated 3 days ago

Qovaryx
/

qovaryx-options-decoder-full-community