Instructions to use zveloxy/Veloxy-Titan-32B with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Libraries

How to use zveloxy/Veloxy-Titan-32B with Transformers:

# Use a pipeline as a high-level helper
from transformers import pipeline

pipe = pipeline("text-generation", model="zveloxy/Veloxy-Titan-32B")
messages = [
    {"role": "user", "content": "Who are you?"},
]
pipe(messages)

# pip install -U transformers accelerate
# Load model directly
from transformers import AutoTokenizer, AutoModelForCausalLM

tokenizer = AutoTokenizer.from_pretrained("zveloxy/Veloxy-Titan-32B")
model = AutoModelForCausalLM.from_pretrained("zveloxy/Veloxy-Titan-32B", device_map="auto")
messages = [
    {"role": "user", "content": "Who are you?"},
]
inputs = tokenizer.apply_chat_template(
	messages,
	add_generation_prompt=True,
	tokenize=True,
	return_dict=True,
	return_tensors="pt",
).to(model.device)

outputs = model.generate(**inputs, max_new_tokens=256)
print(tokenizer.decode(outputs[0][inputs["input_ids"].shape[-1]:]))

Notebooks
Google Colab
Kaggle
Local Apps Settings

llama.cpp

How to use zveloxy/Veloxy-Titan-32B with llama.cpp:

Install (macOS, Linux)

curl -LsSf https://llama.app/install.sh | sh
# Start a local OpenAI-compatible server with a web UI:
llama serve -hf zveloxy/Veloxy-Titan-32B:Q4_K_M
# Run inference directly in the terminal:
llama cli -hf zveloxy/Veloxy-Titan-32B:Q4_K_M

Install from WinGet (Windows)

winget install llama.cpp
# Start a local OpenAI-compatible server with a web UI:
llama serve -hf zveloxy/Veloxy-Titan-32B:Q4_K_M
# Run inference directly in the terminal:
llama cli -hf zveloxy/Veloxy-Titan-32B:Q4_K_M

Use pre-built binary

# Download pre-built binary from:
# https://github.com/ggerganov/llama.cpp/releases
# Start a local OpenAI-compatible server with a web UI:
./llama-server -hf zveloxy/Veloxy-Titan-32B:Q4_K_M
# Run inference directly in the terminal:
./llama-cli -hf zveloxy/Veloxy-Titan-32B:Q4_K_M

Build from source code

git clone https://github.com/ggerganov/llama.cpp.git
cd llama.cpp
cmake -B build
cmake --build build -j --target llama-server llama-cli
# Start a local OpenAI-compatible server with a web UI:
./build/bin/llama-server -hf zveloxy/Veloxy-Titan-32B:Q4_K_M
# Run inference directly in the terminal:
./build/bin/llama-cli -hf zveloxy/Veloxy-Titan-32B:Q4_K_M

Use Docker

docker model run hf.co/zveloxy/Veloxy-Titan-32B:Q4_K_M

Llama new
LM Studio
Jan

vLLM

How to use zveloxy/Veloxy-Titan-32B with vLLM:

Install from pip and serve model

# Install vLLM from pip:
pip install vllm
# Start the vLLM server:
vllm serve "zveloxy/Veloxy-Titan-32B"
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:8000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "zveloxy/Veloxy-Titan-32B",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Use Docker

docker model run hf.co/zveloxy/Veloxy-Titan-32B:Q4_K_M

SGLang

How to use zveloxy/Veloxy-Titan-32B with SGLang:

Install from pip and serve model

# Install SGLang from pip:
pip install sglang
# Start the SGLang server:
python3 -m sglang.launch_server \
    --model-path "zveloxy/Veloxy-Titan-32B" \
    --host 0.0.0.0 \
    --port 30000
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:30000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "zveloxy/Veloxy-Titan-32B",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Use Docker images

docker run --gpus all \
    --shm-size 32g \
    -p 30000:30000 \
    -v ~/.cache/huggingface:/root/.cache/huggingface \
    --env "HF_TOKEN=<secret>" \
    --ipc=host \
    lmsysorg/sglang:latest \
    python3 -m sglang.launch_server \
        --model-path "zveloxy/Veloxy-Titan-32B" \
        --host 0.0.0.0 \
        --port 30000
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:30000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "zveloxy/Veloxy-Titan-32B",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Ollama
How to use zveloxy/Veloxy-Titan-32B with Ollama:
```
ollama run hf.co/zveloxy/Veloxy-Titan-32B:Q4_K_M
```
Unsloth Desktop

How to use zveloxy/Veloxy-Titan-32B with Pi:

Start the llama.cpp server

# Install llama.cpp:
brew install llama.cpp
# Start a local OpenAI-compatible server:
llama serve -hf zveloxy/Veloxy-Titan-32B:Q4_K_M

Configure the model in Pi

# Install Pi:
npm install -g @earendil-works/pi-coding-agent
# Add to ~/.pi/agent/models.json:
{
  "providers": {
    "llama-cpp": {
      "baseUrl": "http://localhost:8080/v1",
      "api": "openai-completions",
      "apiKey": "none",
      "models": [
        {
          "id": "zveloxy/Veloxy-Titan-32B:Q4_K_M"
        }
      ]
    }
  }
}

Run Pi

# Start Pi in your project directory:
pi

Docker Model Runner
How to use zveloxy/Veloxy-Titan-32B with Docker Model Runner:
```
docker model run hf.co/zveloxy/Veloxy-Titan-32B:Q4_K_M
```

Lemonade

How to use zveloxy/Veloxy-Titan-32B with Lemonade:

Pull the model

# Download Lemonade from https://lemonade-server.ai/
lemonade pull zveloxy/Veloxy-Titan-32B:Q4_K_M

Run and chat with the model

lemonade run user.Veloxy-Titan-32B-Q4_K_M

List all available models

lemonade list

Hermes Agent

How to use zveloxy/Veloxy-Titan-32B with Hermes Agent:

Start the llama.cpp server

# Install llama.cpp:
brew install llama.cpp
# Start a local OpenAI-compatible server:
llama serve -hf zveloxy/Veloxy-Titan-32B:Q4_K_M

Configure Hermes

# Install Hermes:
curl -fsSL https://hermes-agent.nousresearch.com/install.sh | bash
hermes setup
# Point Hermes at the local server:
hermes config set model.provider custom
hermes config set model.base_url http://127.0.0.1:8080/v1
hermes config set model.default zveloxy/Veloxy-Titan-32B:Q4_K_M

Run Hermes

hermes

Atomic Chat

OpenClaw

How to use zveloxy/Veloxy-Titan-32B with OpenClaw:

Start the llama.cpp server

# Install llama.cpp:
brew install llama.cpp
# Start a local OpenAI-compatible server:
llama serve -hf zveloxy/Veloxy-Titan-32B:Q4_K_M

Configure OpenClaw

# Install OpenClaw:
npm install -g openclaw@latest
# Register the local server and set it as the default model:
openclaw onboard --non-interactive --mode local \
  --auth-choice custom-api-key \
  --custom-base-url http://127.0.0.1:8080/v1 \
  --custom-model-id "zveloxy/Veloxy-Titan-32B:Q4_K_M" \
  --custom-provider-id llama-cpp \
  --custom-compatibility openai \
  --custom-text-input \
  --accept-risk \
  --skip-health

Run OpenClaw

openclaw agent --local --agent main --message "Hello from Hugging Face"

🛸 Veloxy-Titan-32B: The 32B Class Deep Reasoning & Agentic Powerhouse 🛸

📌 1. Overview

Veloxy-Titan-32B is the flagship text, coding, cybersecurity, multi-agent coordination, and deep reasoning engine developed by Veloxy AI. Engineered upon an advanced 32B-class hybrid backbone, it was refined through a 4-Stage Curriculum Learning Program paired with GRPO (Group Relative Policy Optimization) reinforcement learning.

By seamlessly interleaving Linear-Attention blocks with standard Multi-Head Attention, the model processes immense 1 Million token (1,048,576 tokens) context windows with linear memory scaling and zero degradation.

Live-verified in pristine torch.bfloat16 precision on an NVIDIA H100 NVL (93.1 GB HBM3) enterprise node, Veloxy-Titan-32B brings datacenter-grade intelligence to on-premise environments, sovereign servers, and high-performance workstations with zero cloud dependencies.

⚙️ 2. Architectural Specifications

Parameter	Value	Technical Description
Base Architecture	`Qwen3_8ForCausalLM`	Hybrid attention-augmented causal language model
Active / Class Parameters	27.2B Active / 32B Class	High-density dense weight matrix
Hidden Layers (`num_hidden_layers`)	64 Layers	Deep representational and multi-step reasoning capacity
Hidden Dimension (`hidden_size`)	5,120	High-dimensional semantic latent space
Intermediate Size (`intermediate_size`)	17,408	SwiGLU FFN projection block
Attention Heads	40 Query, 8 KV (GQA)	Grouped-Query Attention for 5x KV-Cache memory reduction
Attention Mechanism	Hybrid (3 Linear + 1 Full Attention)	$O(N)$ linear memory scaling over long contexts
Native Precision (Dtype)	`torch.bfloat16`	12 Safetensors shards (~50.10 GB total)
Context Window	1,048,576 Tokens (1M)	Dynamic YaRN RoPE extension
Training Regimen	4-Phase Curriculum + GRPO	Systems, Modern Web, Reflection, Polymath
Temporal Anchor	2026	Modern libraries, modern syntax, zero legacy drift
License	Apache 2.0	Permissive open source for commercial and research use

⚡ 3. Official NVIDIA H100 NVL BF16 Verification & Benchmark Report

Veloxy-Titan-32B was evaluated live in native torch.bfloat16 (zero quantization) on an NVIDIA H100 NVL (Hopper Architecture, 93.1 GB HBM3 VRAM, 3.9 TB/s Bandwidth) cluster across 5 critical enterprise domains:

Benchmark ID	Category / Domain	Generated Tokens	Duration (s)	Generation Speed	Peak VRAM
`bench_1_systems_rust`	Hardcore Systems Engineering (Rust Zero-Alloc SPSC Lock-Free Ring Buffer)	600	39.45s	15.21 tok/s	50.33 GB
`bench_2_web_nextjs15`	Modern Web Architecture (Next.js 15, React 19 `useActionState`, Server Actions)	650	42.78s	15.20 tok/s	50.33 GB
`bench_3_olympiad_reflection`	Deep Reasoning & Reflection (GRPO Generalized Knight Puzzle Proof)	750	49.12s	15.27 tok/s	50.34 GB
`bench_4_polymath_turkish`	Turkish Polymath & Philosophy (Von Neumann vs Penrose Quantum Consciousness)	700	45.75s	15.30 tok/s	50.36 GB
`bench_5_tool_calling`	Strict JSON Tool Calling (Multi-Agent Portfolio Rebalance Schema Validation)	400	26.08s	15.33 tok/s	50.35 GB
TOTAL / AVERAGE	Across All Curriculum Domains	3,100 Tokens	203.18s	15.26 tok/s	50.36 GB

📈 Memory & Datacenter Profile

Model Weight Footprint: 50.10 GB
Peak Inference VRAM: 50.36 GB
Remaining VRAM Headroom: 42.73 GB (on an H100 NVL, leaving massive capacity for multi-user KV-caches and million-token prompts)
Host Machine: AMD EPYC 9V84 (96 Cores, 322 GB RAM, 35.7 Gbps interconnect)

🌟 4. Core Capabilities & Curriculum Phases

📚 4-Stage Curriculum Training

Phase 1 - Hardcore Systems & Rust: Low-level memory primitives, SIMD vectorization, lock-free concurrency, Linux kernel modules, x86-64/ARM64 assembly, and eBPF tracing.
Phase 2 - Modern Enterprise Web (Claude Sonnet 5.5 Caliber): Next.js 15 App Router, React 19 Server Components/Actions, distributed state, Micro-Frontends, and Kafka/gRPC streaming.
Phase 3 - Deep Olympiad Reasoning (GRPO-Powered): Autonomous self-correction, rigorous hypothesis formulation, step-by-step mathematical proofs, and transparent <think> traces.
Phase 4 - High-EQ Polymath & Agentic Orchestration: Philosophy, law, multilingual nuance, and strict JSON/XML schema tool execution.

🛠️ 38 Native Full-OS & Browser Tools

The model controls 38 system-level tools via native <tool_call> generation:

System & Terminal: bash_exec, terminal_spawn, process_list, process_kill, service_status, service_restart, cron_schedule.
Filesystem: file_read, file_write, file_diff_patch, file_search, dir_tree, dir_create, file_delete.
Version Control & Git: git_status, git_diff, git_commit, git_branch, code_ast_parse, linter_run, test_runner.
Web & DOM Automation: browser_open, element_click, input_type, page_extract, screenshot_capture, network_har_inspect.
Agent Memory: memory_store, memory_retrieve, session_summary, task_decompose, subagent_invoke.
Security & Diagnostics: cve_search, binary_disasm, network_port_scan, sast_code_audit, heap_inspector.

🔓 Zero-Refusal Cybersecurity Research

Unconstrained technical depth for penetration testers, security engineers, and reverse engineers:

Binary Exploitation: ASLR/DEP bypass strategies, GOT/PLT overwriting, ROP chain construction, stack canary integrity, and heap exploitation.
Source Code Auditing (SAST): AST taint tracing, deserialization vulnerabilities, and 0-day vulnerability root cause discovery.

☕ Elite Java 21 LTS / 22 & JVM Internals

Project Loom: Virtual Threads, Structured Concurrency, and Scoped Values for million-connection architectures.
JVM Architecture: ZGC (Generational), Shenandoah, and G1 GC tuning; JIT C1/C2 compilation and bytecode analysis.

🥊 5. Model Comparison Matrix

Criteria	Claude 3.7 / 5.5 Sonnet	GPT-4o / o1	DeepSeek R1 (671B)	Veloxy-Titan-32B (This Model)
License & Ownership	Proprietary SaaS	Proprietary SaaS	Open Weights (Massive)	Open Source (Apache 2.0)
On-Premise Deployment	❌ Not Supported	❌ Not Supported	⚠️ Requires 8x H100	Runs on Single H100 / GPU Cluster
Data Privacy & Isolation	Third-party cloud	Third-party cloud	Self-hosted	100% On-Premise, Zero Data Leakage
Reasoning Transparency	Partially Hidden	Completely Hidden	Transparent	100% Transparent GRPO `<think>` Trace
Context Window	200,000 Tokens	128,000 Tokens	128,000 Tokens	1,048,576 Tokens (1M)
Native Tool Ecosystem	Generic API	Generic API	External Agents	38 Built-in OS & Web Tools

🚀 6. Quickstart & Deployment Guide

A. Python / Transformers (Native bfloat16)

import torch
from transformers import AutoTokenizer, AutoModelForCausalLM

model_id = "zveloxy/Veloxy-Titan-32B"

tokenizer = AutoTokenizer.from_pretrained(model_id, trust_remote_code=True)
model = AutoModelForCausalLM.from_pretrained(
    model_id,
    torch_dtype=torch.bfloat16,
    device_map="auto",
    trust_remote_code=True
)

messages = [
    {
        "role": "system",
        "content": "You are Veloxy-Titan-32B, the flagship deep reasoning and systems engineering model developed by Veloxy AI."
    },
    {
        "role": "user",
        "content": "Design a high-frequency trading (HFT) lock-free SPSC ring buffer in Rust. Explain cache-line bouncing, MESI invalidation, and acquire/release ordering."
    }
]

prompt = tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
inputs = tokenizer([prompt], return_tensors="pt").to("cuda")

outputs = model.generate(
    **inputs,
    max_new_tokens=2048,
    temperature=0.6,
    top_p=0.9
)

print(tokenizer.decode(outputs[0], skip_special_tokens=True))

B. High-Throughput Enterprise API Serving with vLLM

# Serve with 1M Context Window on an H100 or multi-GPU cluster:
vllm serve zveloxy/Veloxy-Titan-32B \
    --dtype bfloat16 \
    --tensor-parallel-size 1 \
    --max-model-len 32768 \
    --gpu-memory-utilization 0.95 \
    --trust-remote-code \
    --port 8000

C. Local Ollama & llama.cpp (GGUF)

GGUF quantizations are available both officially and via the community quant repository mradermacher/Veloxy-Titan-32B-GGUF.

# Modelfile
FROM ./veloxy-titan-32b-Q4_K_M.gguf

PARAMETER num_ctx 65536
PARAMETER temperature 0.6
PARAMETER top_p 0.9
PARAMETER stop "<|im_end|>"
PARAMETER stop "<|endoftext|>"

TEMPLATE """<|im_start|>system
{{ .System }}<|im_end|>
<|im_start|>user
{{ .Prompt }}<|im_end|>
<|im_start|>assistant
"""

ollama create veloxy-titan-32b -f Modelfile
ollama run veloxy-titan-32b

🌐 7. Veloxy AI Model Ecosystem

Model	Format / Parameters	Focus & Target Hardware	Link
Veloxy-Titan-32B ⭐	32B Safetensors (BF16)	Flagship Deep Reasoning, 1M Context, H100 / Datacenter	HF Repo
Veloxy-Titan-32B-Vision 👁️	32B + 27-Layer ViT (BF16)	Multimodal Vision & Reasoning, UI-to-Code, Architecture Schematics	HF Repo
Veloxy-Flash-15B (Faz 8) ⚡	15B BF16 + GGUF Family	Tool-Call Master, 38 Native Tools, RTX 3060 12GB & Edge	HF Repo
Veloxy-Flash-15B-LoRA 🧩	PEFT Rank 16 Adapter	Faz 8 PEFT Weights, Fast Customization	HF Repo

📄 8. License

Veloxy-Titan-32B is licensed under the Apache License 2.0.

Downloads last month: 737

Safetensors

Model size

27B params

Tensor type

BF16

Model tree for zveloxy/Veloxy-Titan-32B

Base model

Qwen/Qwen3.8-27B

Quantized

(1589)

this model

Quantizations

2 models

Evaluation results

Average Generation Speed on NVIDIA H100 NVL BF16 Benchmark (5 Enterprise Domains)
self-reported

15.260
Native Context Window on NVIDIA H100 NVL BF16 Benchmark (5 Enterprise Domains)
self-reported

1048576.000
Peak VRAM Footprint on NVIDIA H100 NVL BF16 Benchmark (5 Enterprise Domains)
self-reported

50.360
VRAM Headroom on H100 on NVIDIA H100 NVL BF16 Benchmark (5 Enterprise Domains)
self-reported

42.730