Syntiox-1.0-Flash (3.7B)
Hugging Face | GitHub | Launch Blog | Documentation
License: Apache 2.0 | Authors: Syntiox Research Team & Developer Community
Syntiox-1.0-Flash is a 3.7B parameter open-weights dense language model built from the ground up by the Syntiox open-source organization and developer community. Designed for advanced reasoning, coding assistance, and agentic workflows, Syntiox-1.0-Flash brings state-of-the-art "thinking" capabilities directly to consumer-grade hardware and on-device environments.
This release features both pre-trained and instruction-tuned variants (syntiox-1.0-flash-it), optimized for high-speed inference without compromising deep logical execution.
Key Capability & Architectural Advancements
- Native "Flash-Thinking" Mode: Features a built-in step-by-step internal reasoning mechanism, allowing the model to decompose complex math, logic, and coding problems before generating the final answer.
- Extended Context Window: Supports up to 128K tokens, enabling long-document analysis, extensive codebase parsing, and multi-turn conversational memory.
- Optimized for Consumer Hardware & On-Device: With a compact 3.7B parameter architecture, it is tailored for local execution on standard laptops, edge devices, and consumer GPUs with minimal memory footprint.
- Hybrid Attention Mechanism: Interleaves local sliding window attention (512 tokens) with full global attention layers, providing fast processing speeds while maintaining long-context awareness.
- Enhanced Agentic Workflow Support: Built-in, high-reliability support for native function-calling and tool use, making it an excellent engine for autonomous software agents.
Model Overview
| Property | Syntiox-1.0-Flash (Dense) |
|---|---|
| Total Parameters | 3.7B (Effective parameters) |
| Layers | 32 |
| Sliding Window Size | 512 tokens |
| Context Length | 128,000 (128K) tokens |
| Vocabulary Size | 262,144 (262K) tokens |
| Supported Modalities | Text (Inputs & Outputs) |
| Position Embeddings | Proportional RoPE (p-RoPE) |
Benchmark Results
Syntiox-1.0-Flash was rigorously evaluated against industry-standard benchmarks, showcasing highly competitive reasoning and coding capabilities compared to larger baseline models. (Results listed are for the Instruction-Tuned variant).
| Benchmark | Syntiox-1.0-Flash (3.7B) | Baseline Model A (7B) | Baseline Model B (3B Class) |
|---|---|---|---|
| MMLU Pro | 71.2% | 68.5% | 58.2% |
| AIME 2026 (No Tools) | 44.8% | 35.0% | 21.4% |
| LiveCodeBench v6 | 54.3% | 46.2% | 31.0% |
| GPQA Diamond | 56.1% | 44.5% | 35.2% |
| BigBench Extra Hard | 36.5% | 31.2% | 19.8% |
Getting Started
You can deploy and run Syntiox-1.0-Flash using the standard Hugging Face transformers library.
Installation
Ensure your environment is up to date:
pip install -U transformers torch accelerate
Basic Usage Inference
from transformers import AutoTokenizer, AutoModelForCausalLM
import torch
MODEL_ID = "syntiox/syntiox-1.0-flash-it"
# Load tokenizer and model
tokenizer = AutoTokenizer.from_pretrained(MODEL_ID)
model = AutoModelForCausalLM.from_pretrained(
MODEL_ID,
torch_dtype=torch.bfloat16,
device_map="auto"
)
# Structure prompt using Native System Prompt Support
messages = [
{"role": "system", "content": "You are Syntiox AI V1, a helpful and precise assistant."},
{"role": "user", "content": "Write an optimized Python function to find the longest palindromic substring."}
]
# Apply chat template (To enable reasoning/thinking mode, keep enable_thinking=True if supported)
text = tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
inputs = tokenizer(text, return_tensors="pt").to(model.device)
input_len = inputs["input_ids"].shape[-1]
# Generate Response
outputs = model.generate(**inputs, max_new_tokens=2048, temperature=0.7, top_p=0.95)
response = tokenizer.decode(outputs[0][input_len:], skip_special_tokens=True)
print(response)