GLM-5.2 is live. #1 throughput on OpenRouter, pay-per-token on FriendliAI. Try it today ➜
Cynix7
Available on FriendliAI
Run this model inference on single tenant GPU with unmatched speed and reliability at scale.
Model Details
Model Provider
Cynix7
Model Tree
tjarvis91/vfai-x-3.5-9b-options
Input Modalities
Output Modalities
Supported Functionality
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 sidecar AI status, removes the legacy vLLM/Docker/WSL waiting path when Qovaryx is healthy, repairs the bundled CPU Torch runtime, restores Electron V8 startup snapshots, preserves user-entered Tradier/API keys, prevents balance/positions blinking during transient Tradier refresh failures, repairs stale Qovaryx runtime downloads, and keeps the public build keyless.
Model/runtime source: Qovaryx/qovaryx-options-decoder-full-community.
Auto-update pointer: updates/latest.json
This is the same 2.5.3 public hotfix listed above. It ships no API keys or broker credentials. Existing installs pull this update from the Hugging Face public update channel.
🚀 New flagship: Qovaryx Options Decoder — Full Community Runtime. The latest, most capable Qovaryx release is live as a single drop-in package: six functional HGB specialists + eight vaulted torch heads in one runtime. 15-of-15 internal benchmark cells closed at strict bootstrap CI lower bound. Drop-in replacement for FrankenB / V3.7 / Qwen-VPA. Sub-millisecond inference. Offline. No license email required. 👉 Qovaryx/qovaryx-options-decoder-full-community
💬 Join the community. Discord: https://discord.gg/PtuHZDv5ju — builders training their own trading/finance models. Engineering, no signals. Get install help, share work, follow the Qovaryx research devlog. Try the deployed Q-Chat router live via
/qchat ask. Ko-fi: https://ko-fi.com/tjarvis91 — every coffee literally buys GPU time for the next training cycle.
The current shipping desktop app (VFAi-X Qovaryx 2.5.3, version 3.5v31-2.5.3-qovaryx-public, links above) no longer uses the Franken-B 9B vision-language model as its inference engine. It uses the Qovaryx Options Decoder community runtime (Qovaryx/qovaryx-options-decoder-full-community) which runs on CPU, has no GPU requirement, and serves sub-millisecond inference. See the New flagship banner above and the Qovaryx card for what currently powers your installs.
Everything below this point is the historical Franken-B model lineage that this repo originally documented. It is preserved because the Franken-B research (composition-lock, last-3 LoRA, vision-fusion +106%) is part of the public record of how the VFAi-X line evolved, and the weights are still hosted here for researchers who want to study or reproduce the 9B vision-language path. It is not what the current desktop app loads at runtime.
The first VFAi-X model to statistically beat the V5.0 baseline on a per-trade bootstrap (penny, p=0.023). Best chart-fusion result in the entire VFAi-X lineage (+106.76% vision-only on the 200-row real fusion harness). Same overlay won on all three streams over 10,000+ acted trades.
Franken-B is a Qwen3.5-9B vision-language model for US equity options trading, built by stacking three LoRA adapters on Qwen3.5-9B-VL -- V3.7s base → V5.0 LM LoRA → V5.8 last-3 LM LoRA. It reads composite chart images (1D / 1H / 5M) plus structured market context and emits BUY / SELL / HOLD / NO_TRADE with conviction and a risk plan. Optimized for Volume-Price Analysis (VPA), pattern recognition, and chart-text fusion.
Production branch: main (this page) · Recipe: V3.7s → V5.0 → V5.8 last-3 · Base: Qwen/Qwen3.5-9B-VL · Runtime: FP8 on vLLM · Predecessor: tjarvis91/vfai-x-3.5-9b-options (V3.7)
If this work helps you, support continued development: ko-fi.com/tjarvis91 ☕
See what you get before you install. Six tabs, real Tradier-connected paper account, model serving Franken-B locally on consumer GPU. (AI engine was disabled for these screenshots so we didn't compete with an active training run -- the engine connects automatically on first launch.)
Local-first Windows desktop app with vLLM acceleration.
Franken-B flagship model path and served-model lock.
Hybrid stocks/options operation with constrained penny mode enabled by default.
Tradier broker bridge for user-supplied paper/live credentials.
Public-safe settings: no bundled broker, Hugging Face, data-provider, or private keys.
Hugging Face auto-update channel for critical hotfixes.
Position sizing: 20% max position size, 8 max positions.
Options overlay: 15x leverage assumption, -5% hard stop, 0.5% theta drag.
Signal gates: 70 conviction floor, 1.0 relative-volume floor, 0.70 profit-score floor.
Fusion path: vision, fusion, gems, chart analysis, and position-manager vision remain enabled.
Thinking mode remains off for this public runtime profile.
The installer is about 1.1 GB and does not bundle the model weights. On first launch it downloads the current Franken-B model files from Hugging Face, verifies the 6 safetensor shards and processor/tokenizer files, then starts vLLM.
If the model repo is ever private or gated, add a Hugging Face read token in Settings or set HF_TOKEN before first launch. Public builds ship no secrets by design.
Independent training runs cost real GPU time. Support is optional and never gates the app: ko-fi.com/tjarvis91
Same overlay sweep (min_conv=70, lev=15, hard_stop=-5%, theta_drag=0.5, position_size=20%, max_pos=8) applied identically to both models, same streams, same dates:
| Metric | V3.7 (prior flagship) | Franken-B (this model) | Delta |
|---|---|---|---|
| 2yr return | +1,432% | +8,245% | +6,813 pp |
| 2yr profit factor | 2.99 | 4.51 | +50.8% |
| Penny return | +12,128% | +35,155% | +23,027 pp |
| Penny profit factor | 26.77 |
Aggregate acted trades: n=10,396 (n=5,030 + 2,822 + 2,544 per stream). Same overlay won every stream -- no overfitting to one regime.
These are simulated overlay backtests against locked answer keys we have audited and found to be selection-biased (286,000 directional rows, zero losers by construction). They are upper bounds -- treat them as relative to other models tested under the same conditions, not as forward returns. Walk-forward validation is queued.
When restricted to the same (symbol, signal_date) pairs both models chose to trade, with identical overlay:
| Stream | V3.7s | Franken-B | Delta |
|---|---|---|---|
| 2yr | n=749, +889%, PF=3.55 | n=749, +933%, PF=3.75 | +44 pp, +0.20 PF |
| penny | n=1,253, +8,734%, PF=9.73 | n=1,246, +9,422%, PF=10.56 | +688 pp, +0.83 PF |
| 180d | n=966, +593%, PF=1.74 | n=950, +604%, PF=1.77 | +11 pp, +0.03 PF |
Franken-B picks better on every stream when restricted to the same trade opportunities. This is the most defensible comparison -- the universe is held fixed, so selection bias cancels out.
markdown
Qwen3.5-9B-VL (base)→ V3.7s merged base (text + fusion sidecar, full LM layers)→ V5.0 LM LoRA (abstention + direction repair, all LM layers q/k/v/o + MLP)→ V5.8 last-3 LM LoRA (layers 29/30/31 only, r=64, alpha=128)
V5.8 is not a vision adapter. It only touches model.language_model.layers.29/30/31. Its chart-fusion edge comes from last-layer language-side arbitration over chart-derived tokens passed from the visual encoder. That is the discovery that drives the +106.76% vision result.
The order matters -- every alternative we tested regressed:
| Recipe attempted | Outcome |
|---|---|
V3.7s → V5.8 (skip V5.0) | take rate 87.5%, prose drift, rejected |
V3.7s → V6 → V5.8 | 7/8 valid, schema miss, rejected |
V3.7s → V5.8 → V6 | same schema miss reversed |
V3.7s → V5.0 → V5.8 → V6 | 8/8 strict, but repetition/malformed output |
V5.5 → V5.0 (double-V5.0) | take 50%, over-abstains |
V5.0 sandwiched between V3.7s and V5.8 stabilizes format before V5.8 applies its directional pressure. We call this composition-lock -- recipe behavior depends on order in a way that cannot be flattened into a single repair adapter.
| Metric | Franken-B | V3.7s | Honest read |
|---|---|---|---|
| brutal action_pass | 90.9% | 100% | V3.7s wins 3 brutal cases (open lane: v27_veto stack candidate) |
| app_integrated | 99.0% | 99.75% | tied for practical purposes |
| today_holdout direction | 48.0% | 44.0% | Franken-B ties V3.5 baseline |
| today_holdout BUY/SELL | 48/2 âš |
The trade-off is real and disclosed: Franken-B loses 8 pp of corpus average and 6.7 pp of adversarial in exchange for +62.77 pp vision, +35.26 pp fused, a balanced action mix, today_holdout matching V3.5 baseline, and the only bootstrap PASS vs V5.0 in the model family. We chose to ship Franken-B because the trade-off makes sense for the runtime profile; the corrector that would have fixed the corpus regressions cost more than it gained on every other dimension.
After the tournament we trained four corrector LoRAs (v3, v4, v5, v58R) on a 109-row law-crystal corpus targeting the boundary failures. Each one improved the strict suite at the cost of vision:
| Configuration | Strict overall | Vision | Fused | Decision |
|---|---|---|---|---|
| Franken-B base (shipped) | 72.16% | +106.76% | +103.32% | PROMOTED |
| + v5 corrector | 81.73% | -21.90% | +30.91% | rejected |
| + v6 corrector | -- | -55.92% | -5.04% | rejected |
The strict-suite gain was concentrated on a few dozen rows. The vision loss was a structural change across the whole input distribution. Sample size won. Composition-lock again: chart-fusion behavior is held jointly by V3.7s + V5.0 + V5.8 and any additional weight-level repair destabilizes it.
We publish the corrector failures because the composition-lock finding is the most ML-interesting result in the work, and it is only credible if the failed iterations are visible.
For research or custom inference outside the desktop app:
python
from transformers import AutoModelForImageTextToText, AutoProcessorimport torchmodel = AutoModelForImageTextToText.from_pretrained("tjarvis91/vfaix-vpa-options-trader",torch_dtype=torch.bfloat16,device_map="auto",trust_remote_code=True,)processor = AutoProcessor.from_pretrained("tjarvis91/vfaix-vpa-options-trader",trust_remote_code=True,)
Or with vLLM for FP8 production serving (what the desktop app does):
bash
vllm serve tjarvis91/vfaix-vpa-options-trader \--served-model-name vfai-qwen35-9b-frankenB-v2.2 \--max-model-len 4096 \--gpu-memory-utilization 0.85 \--quantization fp8 \--trust-remote-code
Expected output schema:
text
ACTION: BUYConviction: 73%Reason: Confirmed VPA direction plus clean retest, volume expansion, defined risk.Risk plan: Buy with defined risk at the failed retest level and exit if that level breaks.
The model is designed to choose between BUY / SELL / HOLD / NO_TRADE based on prompt schema, current position state, market conditions, and evidence quality.
| Open item | Severity | What we're doing |
|---|---|---|
| Today_holdout 96% BUY bias | medium | v58R-balanced fusion corpus (rebalance BUY/SELL anchors before retrain) |
| Corpus adversarial -16.7pp vs V6 | medium | v27_veto adapter stack experiment queued |
| Brutal action_pass -9.1pp vs V3.7s | small | Same v27_veto lane |
| Walk-forward on fresh distribution-realistic hold-out | high (process) | Quarterly split queued; we will publish whichever way it lands |
| Selection-biased legacy answer keys |
The desktop app's V2.2 production overlay is locked while these complete.
The legacy repo tjarvis91/vfai-x-3.5-9b-options hosts two opt-in branches for research/ensembling:
V3.7 (prior flagship): --revision v3.7
V6 Sniper (companion, abstains on 98% of opportunities, highest PF in the family -- confirmation filter, not a primary emitter): --revision vfai-x-sniper-options
Franken-B is the installer default. The other two are for operators who want to roll back, ensemble, or research.
Public audits and report artifacts (paths on the developer machine -- request access for the full bundle):
FRANKEN_B_FLAGSHIP_REPORT_20260520.md -- original flagship promotion report
FINAL_DECISION_FRANKEN_B_FLAGSHIP_20260522.md -- corrector lane verdicts
FRANKEN_B_PRODUCTION_DEPLOYMENT_20260522.md -- production overlay lock + V2.2 settings
frankenB_apples_to_v37s.json -- matched-pair comparison data
frankenB_vs_v50_bootstrap.json -- bootstrap CI raw data
frankenB_mega_variable_sweep.json -- overlay sweep across 1,200 configs
-- sanity battery raw
Devlog series (HF Discussions; also mirrored on the Qovaryx research repo on GitHub):
Apache 2.0. This is a research artifact. Not financial advice. Simulated returns are not guaranteed forward returns. Options trading can result in total premium loss. Use paper trading first. Past evaluation performance does not guarantee future performance.
VFAi-X -- Independent behavioral-model systems lab. Local-first. Open-source. No billion-dollar infrastructure.
| +18.7% |
| 180d return | +1,593% | +4,711% | +3,118 pp |
| 180d profit factor | 4.75 | 4.95 | +4.2% |
| Vision-only fusion | +43.99% | +106.76% | +62.77 pp |
| Fused (text+chart) | +68.06% | +103.32% | +35.26 pp |
| 2yr BUY/SELL balance | 264 / 1732 (87% SELL-skew) | 1,539 / 1,491 (balanced) | bias fixed |
| 2yr direction accuracy | 39.4% | 40.5% | +1.1 pp |
| Bootstrap CI vs V5.0 (penny) | TIE (p=0.39) | PASS p=0.023 | first PASS in family |
V3.7s → V5.0 → V5.8| 8/8, 75% take, balanced action mix -- winner |
| 25/25 |
| bull-bias is the known open weakness |
| corpus avg (9 packs) | 88.2% | 96.2% | V3.7s wins broad corpus (sample-size-small per pack) |
| corpus adversarial | 83.3% | 90.0% | open lane: targeted V58R challenger in progress |
| fusion vision_only | +106.76% | +43.99% | best in the entire family (n=200 fusion harness) |
| fusion fused (with app gate) | +103.32% | +68.06% | best in the family |
| + v58R (latest, 2026-05-22) | sanity match | +2.55% (only positive corrector) | -- | rejected (adversarial collapsed to 73.3%) |
| high (process) |
| Telemetry-gold streams in development |
frankenB_full_20260520_LIVE_SUITE_REPORT.jsonV5_FRANKENB_2YR_PROFITABILITY.md, V5_FRANKENB_PENNY_PROFITABILITY_50USD.md -- locked production profitability
FRANKENB_ALL_REPORTS_SHIP_AUDIT_20260522.md -- pre-ship audit summary
CODEX_V58_REGRESSION_AUDIT_20260522.md -- V5.8 lineage audit + composition-lock finding