Indian AI Lab 169Pi Turns Qwen Into a Tiny Local Multimodal Model
169Pi's 9B vision-language model runs offline on laptops and phones, hitting 61.20% on SWE-Bench Verified with Apache 2.0 weights.
Takeaways
- 169Pi released Alpie Seed, a 9B multimodal model based on Qwen3.5-9B, Apache 2.0 licensed.
- Reports 61.20% on SWE-Bench Verified, 84.32% on GPQA Diamond, 72.99% on LiveCodeBench v6.
- Ships Q2, Q3, Q4 GGUF builds (3.8 to 5.6 GB) for llama.cpp and Ollama.
- Native 262K context window with thinking and non-thinking modes toggle.
- Trained in 4-bit from day one, runs on phones, CPUs, and consumer GPUs.
- Community is skeptical of benchmarks; independent replication pending.
Indian AI lab 169Pi has released Alpie Seed, a 9-billion-parameter model packaged for local inference through llama.cpp and Ollama. The team advertises text, image, and video support, a 262,144-token context window, and quantized files small enough for laptops and some high-memory phones. Its strongest benchmark results remain vendor-reported and need independent replication.
The repository lists the weights under Apache 2.0, which generally permits commercial modification and distribution when its license and notice requirements are met. Developers planning commercial use should also audit the licenses for the Qwen base model, inference runtime, visual projector, and other bundled components.
Qwen underneath, 169Pi on top
Alpie Seed starts with Qwen3.5-9B and adds post-training for reasoning, instruction following, coding, and visual understanding. According to the source repository, 169Pi used supervised fine-tuning and reinforcement learning. Supervised fine-tuning teaches desired responses from curated examples, while reinforcement learning adjusts behavior using reward signals.
The model provides thinking and non-thinking modes. Thinking mode spends additional generation tokens on multi-step reasoning, while non-thinking mode favors faster responses. Actual latency and output quality depend on the prompt template, quantization, inference backend, and hardware.
Strong scores await outside checks
169Pi reports the following results for its merged FP16 checkpoint:
| Benchmark | Score | What it measures |
|---|---|---|
| MMLU-Pro | 82.54% | Broad knowledge and reasoning |
| GPQA Diamond |