FOUNDATION MODEL SERIES 1.1
Swen 1.1 Model Family:
Sub-2B Edge Intelligence.
Breaking the memory wall with hybrid causal convolutions, grouped-query attention, 128,000 native context tokens, and test-time reasoning.
Flagship Instruct1.17 Billion
Swen-1.1-Instruct
High-Efficiency Compact Language Model for Chat, Code & Agentic Workflows
Pure Math Specialist350 Million
Swen-1-Math
Ultra-Compact Mathematical Specialist with Autonomous Chain-of-Thought
Metacognitive Reasoning1.17 Billion
Swen-1.1-Thinking
Metacognitive Reasoning Engine for Self-Reflection & Backward Induction
Flagship Instruct
Swen-1.1-Instruct
High-Efficiency Compact Language Model for Chat, Code & Agentic Workflows
Params: 1.17 Billion
Weight: ~2.34 GB BF16
HumanEval
66.0% 🥇
Pass@1 Code SynthesisGSM8K
54.0% 🥇
Grade School MathGPQA Diamond
38.0% 🥇
Graduate QAContext
128k
Native Token WindowArchitectural Breakthroughs:
Crushes category: SmolLM2-1.7B (22.6%) and Llama-3.2-1B (25.0%) on HumanEval
Native Agentic Tool Use with built-in <|tool_call_start|> protocol
Multilingual comprehension across 8 major languages (English, Hindi, Spanish, etc.)
Runs comfortably on consumer laptops and edge devices with 4GB RAM
Inference & Prompt Protocol
<|im_start|>system
You are Swen-1.1, an autonomous coding agent by Sorika Labs.<|im_end|>
<|im_start|>user
Write a high-throughput async batch tensor cache in PyTorch.<|im_end|>
<|im_start|>assistant
<|tool_call_start|>
{"tool": "python_eval", "code": "import torch
..."}
<|tool_call_end|>