Alpie Seed 9B: Offline AI Model Setup, Specs and Benchmarks

Last updated: November 11, 2026

Alpie Seed 9B is an open-weight model from 169Pi that runs fully offline on laptops, workstations and some phones. It reads text, images and video, has thinking and non-thinking modes, and lists a 262K-token context window. It is free under Apache 2.0 and installs with one Ollama command.

Contents

What Alpie Seed is and what is new
Specs, builds and where to get it
How to run Alpie Seed locally
Benchmarks compared with Qwen3.5-9B
Speed on real devices
Workflows for power users
Limitations
Frequently asked questions

This guide is based on 169Pi’s Alpie Seed 9B research post, the official Alpie Seed product page, the Hugging Face model card and the Ollama library page.

What Alpie Seed is and what is new

169Pi launched Alpie Seed on November 6, 2026, according to its launch announcement. The company describes it as a 9-billion-parameter hybrid-reasoning vision-language model built “to run across devices, not just data centres.”

The model keeps the Qwen3.5-9B architecture. 169Pi changed its behavior through post-training: a DoRA adapter trained with supervised fine-tuning, then reinforcement learning with GRPO. The research post says the reward system scored seven factors, including correctness, a hallucination penalty and a length penalty.

Compared with its Qwen3.5-9B base, the main changes for users are higher reported scores on reasoning and coding tests, ready-made 2-bit, 3-bit and 4-bit builds, and device testing on phones and CPU-only PCs.

Specs, builds and where to get it

Item	Alpie Seed 9B
Developer	169Pi
Parameters	9B
Base model	Qwen3.5-9B, post-trained with SFT and GRPO
Input	Text, image and video
Modes	Thinking and non-thinking
Context window	262,144 tokens native (Ollama tags list 256K)
Builds	16-bit weights plus 2-bit, 3-bit and 4-bit GGUF
License	Apache 2.0
Where to get it	Hugging Face, Ollama, Kaggle; launch materials also name LM Studio and PocketPal

The Ollama page lists four tags. The default alpie-seed:latest is about 5.6 GB, q2 is 3.8 GB and q3 is 4.6 GB; all three are text only. The vision tag is 6.6 GB and adds image input.

You can also try the model in the browser through the Alpie Playground. 169Pi has opened invite-only access to Alpie Local, a coding agent built on Alpie Seed, with wider access planned later in November.

How to run Alpie Seed locally

These steps follow the Ollama page, the GGUF model card and the main model card:

Install Ollama and run ollama run 169pi/alpie-seed to download and chat with the default 4-bit build.
On a machine with little memory, use ollama run 169pi/alpie-seed:q2 or :q3 instead.
For image questions, run ollama run 169pi/alpie-seed:vision.
To use llama.cpp, download the GGUF files with hf download 169Pi/Alpie_Seed_GGUF, pick the Q2, Q3 or Q4 file, and start llama-cli -m your-file.gguf -cnv. Add -ngl 99 to offload layers to an NVIDIA GPU.
To serve an OpenAI-compatible API, run vllm serve 169Pi/Alpie_seed --max-model-len 262144 --reasoning-parser qwen3. Add --enable-auto-tool-choice --tool-call-parser qwen3_coder for tool calls.
With Transformers 5.2 or later, pass enable_thinking=True or False to the chat template to switch modes.

If you are sizing hardware, our Ollama system requirements answer and LM Studio system requirements answer cover the basics.

Benchmarks compared with Qwen3.5-9B

169Pi reports these results from the FP16 model with extended thinking on. Comparison scores come from other vendors’ published reports and were not re-run by 169Pi, so treat the gaps as indicative.

Benchmark	Alpie Seed 9B	Qwen3.5-9B
GPQA Diamond	84.32	81.7
LiveCodeBench v6	72.99	65.6
SWE-Bench Verified	61.2	60.0
AIME 2026	83.33	69.4
OCRBench	90.30	89.2
MathVision	76.76	78.9

Not every result favors Alpie Seed. On the model card, Qwen3.5-9B is listed higher on MathVision, MMMU-Pro and MMStar. 169Pi also reports evaluations in ten Indian languages, with an overall MMLU Indic score of 73.67%.

Speed on real devices

The research post lists offline tests with quantized builds of about 5.2 to 5.6 GB:

NVIDIA RTX PRO 4500: 110.66 tokens per second
Mac M2 with 16 GB: about 14.1 tokens per second
iPhone 17 Pro: 9.59 tokens per second
Samsung Galaxy Z Flip8: 5.90 tokens per second
Windows, CPU only: 1.95 tokens per second

The product page shows slightly different figures for some devices, such as 106.7 tokens per second on the RTX PRO 4500. Expect your results to vary with runtime, context length and cooling.

Workflows for power users

Private document reading. Use the vision tag to pull text from scans, screenshots and charts without sending files to a cloud service.
Offline coding help. Serve the model with vLLM and point a coding tool at the local OpenAI-compatible endpoint.
Mode switching. Turn thinking on for math and debugging, and off for quick replies to save time and tokens.
Size testing. Compare q2, q3 and the default build on your own prompts before choosing one for a laptop fleet.

For other recent local models, see our Ollama 0.40.15 guide, our Recursive Flint guide and our Bielik 11B v3.1 guide.

Limitations

All benchmark numbers are 169Pi’s own and apply to the FP16 model, not the 2-bit, 3-bit or 4-bit builds.
The 262K context was not tested in 169Pi’s memory profiling, which stopped at 32K tokens.
Handwriting recognition scored lower than printed text, at 76% on the OCRBench category.
The default and q2/q3 Ollama tags are text only; images need the vision tag or a matching vision projector.
169Pi has not published its training data mix.

Frequently asked questions

What is Alpie Seed?

Alpie Seed 9B is an open-weight multimodal model from 169Pi, an AI company based in India. It is built on Qwen3.5-9B and post-trained with supervised fine-tuning and GRPO reinforcement learning. It accepts text, image and video input, supports thinking and non-thinking modes, and is designed to run offline.

Is Alpie Seed free to use commercially?

Yes. 169Pi releases Alpie Seed under the Apache 2.0 license, which allows personal and commercial use subject to the license terms. Running it locally has no per-token charge; your costs are your own hardware and electricity.

How do I run Alpie Seed in Ollama?

Install Ollama and run ollama run 169pi/alpie-seed. The default tag is about 5.6 GB and text only. Use the q2 (3.8 GB) or q3 (4.6 GB) tags for less memory, or the vision tag (6.6 GB) for image input.

What is the Alpie Seed context window?

169Pi lists a native context window of 262,144 tokens. The Ollama tags show 256K. In practice, usable context depends on your runtime settings and memory, and 169Pi’s own memory profiling only went up to 32K tokens.

Can Alpie Seed run on a phone?

Yes, with a quantized build. 169Pi tested it offline on an iPhone 17 Pro at about 9.6 tokens per second and on a Samsung Galaxy Z Flip8 at about 5.9 tokens per second. Its launch materials name the PocketPal app as one way to run it on a phone.

Alpie Seed 9B: Power-User Guide to Offline Setup, Specs and Benchmarks