Bluesky Thread

Jan1-4B: a tiny local model that beats Perplexity Pro

August 12, 2025 View original thread

Jan1-4B: a tiny local model that beats Perplexity Pro

attach any search-related MCP and use vLLM or llama.cpp, or use the Janus app

based on Qwen3 and fine tuned for reasoning

huggingface.co/janhq/Jan-v1...

Bar chart titled “SimpleQA Benchmark Results – Comprehensive accuracy on question answering.” Bars represent accuracy percentages for different models.
From highest to lowest:
• Jan-v1 – 91.1% (highlighted in teal with a gold star).
• Perplexity Pro – 90.6% (blue).
• Qwen3-4B-2-2507 – 86.5%.
• gpt-oss-20B – 86.3% (green).
• Jan-nano-128k – 83.2%.
• Jan-nano – 80.7%.
• Jan-nano (YaRN) – 79.7%.
• Lucy (YaRN) – 78.3%.
• DeepSeek-V3 – 78.2%.
• ChatGPT-4.5 – 62.5%.
• Baseline – 59.2%.
• Gemini-2.5-Pro – 52.9%.
• Claude-3.7-Sonnet – 50.0%.
• o3 – 49.4%.
• Grok-3 – 44.6%.
• o1 – 42.6%.

25 4

ngl the Jan app looks neat.

free & open source. model hub. MCP. coming soon: memory & stateful assistants

jan.ai

jan.ai

Jan: Open source ChatGPT-alternative that runs 100% offline - Jan

Chat with AI without privacy concerns. Jan is an open-source ChatGPT-alternative, running AI models locally on your device.

15 3

More like this