VexelQuant Foundry•4 Flagship Quant Models Live on Hugging Face

Replace 900ms Cloud LLMs with 16ms Local Small Language Models.

VexelQuant trains private, GRPO-aligned 1.5B–3B Small Language Models (safetensors + Q4_K_M.gguf) that execute locally in 16–22ms with 98.3%–100.0% strict JSON accuracy—replacing slow 900ms+ external cloud APIs for quantitative trading, prediction market arbitrage, and sniper bots.

Commission Custom Model ($0 Upfront)Inspect 4 Flagship Models Test live in Private Vault before paying →

Live Studio Telemetry & Verification Matrix

Verified Offline Baseline

Median Local Latency

15.8ms – 22.1ms

vs ~942ms Cloud LLM (58x faster)

Strict JSON Compliance

98.3% – 100.0%

Zero markdown fences • 0 backticks

HF Community Downloads

628+

Hugging Face Verified Weights

Marginal Inference CostLOCAL

$0.00 / 1M calls

Runs on your server or laptop

Deterministic Execution Benchmark

Cloud LLM API vs. VexelQuant Local GGUF Trace

In algorithmic trading, prediction market CLOB arbitrage, and DEX snipers, execution is lost when a cloud LLM takes 900ms or hallucinates markdown fences. Compare the real execution traces below.

Incoming Raw Trading Payload / Event Prompt:Input tokens: ~38

US SEC approves BlackRock spot Solana ETF 19b-4 filing; Cboe BZX listing confirmed for Tuesday.

Frontier Cloud LLM APIGPT-4o / Claude 3.5

942ms

SUCCESS RATE84.2%

API COST$5.00

JSON COMPLIANCEFAILED (FENCES)

RAW UNFILTERED HTTP RESPONSE:

```json
{
  "ticker": "SOL",
  "catalyst_type": "REGULATORY",
  "direction_bias": "LONG",
  "urgency_score": 0.98,
  "confidence": 0.99,
  "summary_50c": "SEC formally grants BlackRock SOL ETF approval"
}
```
Note: This is breaking news from regulatory authorities.

Deterministic Bot Failure Modes:

×Wrapped in ```json markdown fence (causes JSON.parse error)

×Added conversational filler footnote at bottom

×942ms round-trip latency misses high-frequency orderbook queue

VexelQuant Quant SLMQ4_K_M.gguf / safetensors

15.8ms (58x faster)

SCHEMA COMPLIANCE100.0%

LOCAL COST$0.00 / 1M

RAW JSON FENCES0 DETECTED

STRICT PARSED JSON STREAM:

{
  "ticker": "SOL",
  "catalyst_type": "REGULATORY",
  "direction_bias": "LONG",
  "urgency_score": 0.98,
  "confidence": 0.99,
  "summary_50c": "SEC formally grants BlackRock SOL ETF approval"
}

Deterministic Runtime Advantages:

&check;Raw pure JSON output with zero markdown or backticks

&check;15.8ms median local latency on consumer GPU / Apple Silicon

&check;100.0% schema conformity enforced by GRPO reward alignment

Production Artifacts & Weights

4 Flagship Quant Models Live on Hugging Face

Trained on private quant execution datasets with Unsloth QLoRA and GRPO schema reinforcement. Ready to download in raw GGUF and safetensors.

Interactive Model Playground huggingface.co/VexelQuant

VexelQuant/Vexel-Catalyst-1.5B-v1•Qwen2.5-1.5B-Instruct Backbone

Vexel-Catalyst-1.5B-v1

High-frequency crypto breaking news, regulatory filings & Telegram alpha extraction into directional ticker JSON.

Test in Playground Hugging Face

Median Latency15.8ms

TTFT (Time to First Token)8.2ms

Strict Schema Accuracy100.0%

Artifact Formatssafetensors + Q4_K_M.gguf

Real Benchmark Input Payload:

BREAKING: U.S. SEC officially issues formal grant approval for BlackRock spot Solana ETF 19b-4 exchange filing, trading slated for next Tuesday open on Cboe BZX.

Deterministic Extracted Output:

{
  "ticker": "SOL",
  "catalyst_type": "REGULATORY",
  "direction_bias": "LONG",
  "urgency_score": 0.98,
  "confidence": 0.99,
  "summary_50c": "SEC formally grants BlackRock SOL ETF 19b-4 approval"
}

Developer Tooling

Open-Source Execution Infrastructure

Core research libraries powering our 15ms quantized inference pipeline and GRPO schema enforcement algorithms. Hosted on GitHub.

github.com/VexelQuant

Evaluation & Benchmarks

vexel-eval-harness

Async CLI & library benchmarking local GGUF latency (p50/p95/p99) and strict Pydantic v2 schema compliance.

•Sub-millisecond local TTFT & end-to-end percentile tracking (p50, p95, p99)

•Strict zero-markdown & Pydantic v2 validation rate tracking

•Head-to-head comparison against OpenAI/Anthropic cloud endpoints

•Automated markdown benchmark report generation for quantitative teams

Install via Git (GitHub Hosted):

pip install git+https://github.com/VexelQuant/vexel-eval-harness.git

Reinforcement Learning

grpo-schema-rewards

Deterministic Unsloth/TRL GRPOTrainer reward functions penalizing markdown fences, hallucinated keys, and schema drift.

•Zero-tolerance format reward: strict -2.0 penalty for ```json fences or backticks

•Exact enum & numerical range bounded reward functions for financial constraints

•Native Unsloth & TRL GRPOTrainer plug-and-play integration

•Enables 100% deterministic JSON output from 1.5B–3B parameter backbones

Install via Git (GitHub Hosted):

pip install git+https://github.com/VexelQuant/grpo-schema-rewards.git

Trading Runtime Bridge

slm-execution-bridge

Sub-20ms Python & TypeScript HTTP/2 execution client with GBNF grammar constraints for Polymarket CLOB & Solana/Base bots.

•Low-latency HTTP/2 and IPC socket bridge to local llama.cpp / vLLM instances

•Pre-compiled GBNF grammars enforcing token-level regex and JSON AST

•Polymarket py-clob-client & Solana web3.js zero-copy pipeline adapters

•Failsafe timeout routing with microsecond fallback to deterministic rules

Install via Git (GitHub Hosted):

pip install git+https://github.com/VexelQuant/slm-execution-bridge.git

Zero-Risk Commercial Model

The “Test Live in Your Private Vault Before You Pay” Protocol

We never ask quant teams for upfront deposits. You test your custom SLM on your own live unseen trading prompts inside a dedicated browser sandbox before settling on-chain.

1Submit Spec ($0 Upfront)

Provide your Pydantic v2 JSON schema and 5–10 sample raw trading prompts on our intake page. No credit card or upfront deposit required.

224–48h Private Vault GPU

Receive an unlisted /vault/<key> URL connected to an autonomous NVIDIA T4 GPU node to test your model on unseen prompts in real-time.

3On-Chain USDC Unlock

Once satisfied with the 16ms latency and 100% strict JSON accuracy, settle in USDC (Base or Solana) inside the Vault to instantly download raw .gguf weights.

Transparent Flat-Fee Pricing

Commercial Custom Model Tiers

All tiers include dedicated synthetic alignment dataset generation, QLoRA + GRPO training, and private GPU Vault evaluation.

Specialized Task

Single-Schema Task SLM

Custom 2,500-row synthetic dataset + Unsloth QLoRA & GRPO alignment + Private Vault verification + weights delivery in 24–48h.

$250USDC one-time

Dedicated 2,500-row verified dataset synthesized for your specific trading schema

QLoRA fine-tuning + GRPO schema reinforcement on Qwen2.5 1.5B or 3B

Private Client Vault interactive testing sandbox with real-time latency profiling

Exported Q4_K_M.gguf and FP16 safetensors with SHA-256 validation

Zero upfront deposit — test live in the Vault before settling in USDC

Apply for Tier 1 ($250 USDC)

Most Selected by Bot Operators

Full Quantitative Execution Suite

Everything in Tier 1 + custom GBNF grammar file + tailored Python/TypeScript slm-execution-bridge bot adapter + 14-day weight tuning.

$450USDC one-time

Everything in Tier 1 (2,500-row dataset + GRPO training + Vault sandbox)

Custom GBNF grammar file locking grammar at the inference token sampler level

Tailored Python / TypeScript slm-execution-bridge integration for your bot

14-day post-delivery weight recalibration on real live trade execution logs

Direct asynchronous support via private encrypted Telegram channel

Apply for Tier 2 ($450 USDC)

Production Teams

Continuous Weight Refresh Retainer

Bi-weekly LoRA/GRPO weight recalibration on new production trade logs, exchange upgrades, and shifting market regimes.

$100USDC / month

Bi-weekly LoRA/GRPO model recalibration on your newly recorded trading feeds

Continuous schema adaptation when DEX routers or exchange APIs update

Regression testing against historical benchmark suite on every refresh

Direct private Vault instance maintained for continuous validation

Cancel anytime on-chain with zero lock-in

Apply for Retainer ($100 USDC/mo)