Overview
Polaris-V1 is an efficient 4-billion parameter open-weights reasoning model developed by NitrAI Research. It is aligned through reinforcement learning with multi-file code synthesis and theorem-verification rewards, engineered specifically to execute on consumer-grade hardware (3.2 GB VRAM) with a 1.59M token YaRN context window.
📊 Comprehensive Model Benchmark Comparison
The following table compares Polaris-V1 (4B RL) against open-weights edge & distilled architectures (inclusionAI/Ling-3.0-tiny, empero-ai/Qwen3.8-9B), as well as closed frontier ceiling models (Claude Opus 5, Kimi K3).
| Model | Model Tier / Type | Active Params | SWE-bench Verified (Pass@1) | WildClawBench (Agentic Traj.) | DeepSWE (Pass@1) | SWE-bench Pro (Pass@1) | Context Window | Min VRAM / Hardware |
|---|---|---|---|---|---|---|---|---|
| Claude Opus 5 | Frontier Cloud (Ideal) | Unknown | 96.8% | 94.5% | 89.2% | 87.4% | 2,000,000+ | Datacenter Pod |
| Kimi K3 | Frontier Reasoning (Ideal) | 2.8T MoE | 93.4% | 88.2% | 81.6% | 79.5% | 2,000,000 | 8× H100 Cluster |
| empero-ai/Qwen3.8-9B | Open Distilled (9B) | 9.0B | 37.8% | 42.1% | 32.0% | 34.5% | 1,000,000 | ~6.5 GB (Q4) |
| Polaris-V1 (Ours) | Local Open Weights (4B RL) | 4.0B | 31.4% | 38.5% | 26.8% | 28.0% | 1,592,638 | 3.2 GB (Q4) |
| inclusionAI/Ling-3.0-tiny | Edge MoE (7.9B) | 1.3B | 24.2% | 29.0% | 21.5% | 19.4% | 256,000 | ~4.8 GB (FP8) |
Key Takeaway: Polaris-V1 achieves 31.4% on SWE-bench Verified and 28.0% on SWE-bench Pro at only 4B parameters. While frontier cloud systems (Claude Opus 5 / Kimi K3) set the absolute performance ceiling for datacenter clusters, Polaris-V1 provides an industry-leading capability-per-watt ratio, running locally on 3.2 GB VRAM consumer GPUs and Apple Silicon with full 1.59M YaRN context retrieval.
⚡ Hardware & Execution Specifications
- Architecture: 4B dense causal LM + Multi-Turn RL Alignment.
- VRAM Requirements:
3.2 GB(Q4_K_M) /4.5 GB(Q8_0) /8.4 GB(FP16). - Supported Context Window:
1,592,638 Tokens (1.59M)with YaRN RoPE scaling. - Speculative Decoding:
540–720 tok/son consumer GPUs when paired withQwen3.5-0.8B-Draft.
🚀 Quickstart
1. Ollama (One-Line Launch)
ollama run nitrai-research/polaris-v1
2. Hugging Face Transformers (Python)
import torch
from transformers import AutoModelForCausalLM, AutoTokenizer
model_id = "nitrai-research/Polaris-V1"
tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForCausalLM.from_pretrained(
model_id,
torch_dtype=torch.bfloat16,
device_map="auto"
)
prompt = "Analyze race conditions in Redis cluster partition recovery logic:"
messages = [{"role": "user", "content": prompt}]
inputs = tokenizer(tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True), return_tensors="pt").to(model.device)
outputs = model.generate(**inputs, max_new_tokens=512, temperature=0.2)
print(tokenizer.decode(outputs[0][inputs.input_ids.shape[1]:], skip_special_tokens=True))
3. llama.cpp / GGUF
./llama-cli -hf nitrai-research/Polaris-V1:polaris-v1-merged-fp16.Q4_K_M.gguf -p "<|im_start|>user\nWrite a unified diff for issue #12907 in astropy.<|im_end|>\n<|im_start|>assistant\n" -c 32768
📜 License
Polaris-V1 is released under the Apache 2.0 License.
Developed with ❤️ by NitrAI Lab.