What is Alpie Seed 9B?
Alpie Seed is a 9-billion-parameter model that reads text, images and video, built by 169Pi AI by post-training Qwen3.5-9B with supervised fine-tuning and GRPO reinforcement learning. The architecture is unchanged from Qwen3.5-9B; the work is in the training, aimed at reasoning, instruction following, coding and Indian languages.
It has a 262K-token context window and ships as full-precision weights plus 2-, 3- and 4-bit GGUF builds. On Ollama the text model is 5.6 GB, the 2-bit build 3.8 GB and the vision build 6.6 GB, and 169Pi shows it running fully offline on a phone. It is licensed Apache 2.0.
Is Alpie Seed 9B worth it?
Alpie Seed is a post-trained version of Qwen3.5-9B rather than a new architecture, and 169Pi says so plainly. The value has to come from the training: supervised fine-tuning with a DoRA adapter, then GRPO reinforcement learning, aimed at reasoning, coding, visual understanding and Indic languages including Hindi, Bengali, Tamil, Telugu and Marathi.
The packaging is good. Full weights sit beside 2-, 3- and 4-bit GGUF files, Ollama tags range from 3.8 GB to 6.6 GB with the vision build, and the 262K context window carries over from the base. A demo shows it answering offline on a phone through PocketPal.
The question is margin. The benchmark table compares against Qwen3.5-9B and gpt-oss-20b using scores taken from their model cards, with Alpie run with extended thinking and up to 65,536 output tokens, so the comparison is not like for like. Scores apply only to the full-precision model, not the quantised builds most people will run. It has about 1,000 downloads and 28 likes on Hugging Face, so outside testing is thin.
Where it earns its place
A developer of a Hindi and Tamil study app
Answer students' questions offline on mid-range phones
Ship the 3-bit GGUF build and run it on-device; check answers in each language against the base Qwen3.5-9B before choosing.
A small team prototyping document Q&A
Ask questions about scanned forms and photos locally
Run ollama run 169pi/alpie-seed:vision on a laptop; the 6.6 GB build takes images as input.
Before you commit
- Setup
- ollama run 169pi/alpie-seed, or download weights or GGUF files from Hugging Face or Kaggle.
- Integrations
- Ollama, llama.cpp-compatible GGUF runners, PocketPal on mobile, Transformers.
- Data and privacy
- Runs locally; Apache 2.0.
- Cost at scale
- Free weights; a 9B model runs on a single consumer GPU or a recent laptop.
- Limits
- Benchmarks for full precision only; quantised builds will score lower; base-model limits of Qwen3.5-9B apply.
What’s good
- Targeted post-training for Indian languages
- Text, image and video input with 262K context
- GGUF builds from 3.8 GB; runs offline on a phone
- Apache 2.0
What’s not
- Same architecture as Qwen3.5-9B; gains rest on 169Pi's benchmarks
- Comparison scores taken from other models' cards
- Benchmarks cover only the full-precision weights
- Little independent testing so far
Best for
Developers building offline or Indic-language apps who want a small multimodal model and will test it against Qwen3.5-9B themselves.
Alpie Seed 9B: common questions
- Is Alpie Seed 9B worth it?
- A sensible fine-tune of Qwen3.5-9B with real effort on Indian languages and on-device builds, but it has to prove it beats the model it started from. We rate it 3.7 out of 5.
- How much does Alpie Seed 9B cost?
- Free open weights under Apache 2.0.
- What are the downsides of Alpie Seed 9B?
- Same architecture as Qwen3.5-9B; gains rest on 169Pi's benchmarks; comparison scores taken from other models' cards; benchmarks cover only the full-precision weights; little independent testing so far.
- Who is Alpie Seed 9B best for?
- Developers building offline or Indic-language apps who want a small multimodal model and will test it against Qwen3.5-9B themselves.
- How much does Alpie Seed 9B cost for a team or in production?
- Free weights; a 9B model runs on a single consumer GPU or a recent laptop.
- Is Alpie Seed 9B safe for company data?
- Runs locally; Apache 2.0.
Answers checked Nov 15, 2026.