Imagine you want an AI model to read a confidential contract, a scanned invoice or a long internal report, and you would rather not send any of it to a cloud service. That is the problem Alpie Seed 9B, a new model from the Indian lab 169Pi AI, is designed for: a reasoning model small enough to run on a laptop, a workstation and even a phone, without an internet connection.
What 169Pi actually released
Alpie Seed 9B is an open-weight vision-language model with 9 billion parameters. “Open-weight” means the trained model files themselves are published, so you can download and run them on your own hardware. “Vision-language” means it accepts images, and according to 169Pi also video, alongside text. It is released under the Apache 2.0 licence, which allows personal and commercial use under the licence terms.
The model is not built from scratch. Its Hugging Face model card says it uses the Qwen3.5-9B architecture and was post-trained in two stages: supervised fine-tuning with a DoRA adapter, followed by reinforcement learning with GRPO (Group Relative Policy Optimization), a method that rewards better answers among several sampled ones. The result supports a “thinking” mode, where the model works through a problem step by step before answering, and a faster non-thinking mode.
You can run it through Ollama (ollama run 169pi/alpie-seed, with q2, q3 and vision tags), through GGUF files for tools like llama.cpp and LM Studio, or through Transformers, vLLM and SGLang on stronger hardware. One detail matters for image work: on Ollama, only the vision tag accepts images; the other tags are listed as text-only.
How strong are the benchmark numbers?
169Pi reports 84.32 percent on GPQA Diamond (graduate-level science questions), 82.54 percent on MMLU-Pro, 72.99 percent on LiveCodeBench v6, 61.20 percent on SWE-Bench Verified (real software bug fixes, measured with an agent scaffold) and 90.30 percent on OCRBench, a document and text-reading test. For comparison, the model card lists Qwen3.5-9B, its own base, at 81.7, 82.5, 65.6 and 60.0 on the first four.
These are developer-reported results with extended thinking switched on. The card itself says the scores for other models were taken from their model cards and reports, not re-run in the same test setup, so cross-model comparisons are indicative only. The SWE-Bench figure also depends on the agent scaffold used around the model.
Start with the job you need done, because the gains are not uniform. On text reasoning and coding, Seed sits clearly above its Qwen base in 169Pi’s table. On several vision tests it does not: MMMU-Pro (68.70 versus 70.1), MathVision (76.76 versus 78.9) and MMStar (78.87 versus 79.7) are slightly lower. If your main use is solving maths from images, the post-training has not obviously helped; if it is reasoning, coding or reading documents, the reported numbers point the other way. 169Pi also lists its own weak spots: 76 percent on handwriting recognition, 66 percent on handwritten maths expressions and 13.24 percent on Humanity’s Last Exam.
Does it really run on a laptop or phone?
169Pi’s research page lists offline tests with quantized builds of around 5.2 to 5.6 GB. Quantization stores the weights with fewer bits, which shrinks the model at some cost in precision; 169Pi offers 2-, 3-, 4- and 16-bit builds.
The reported generation speeds are about 110.66 tokens per second on an NVIDIA RTX PRO 4500 workstation GPU, about 14.1 on a Mac M2 with 16 GB, 9.59 on an iPhone 17 Pro, 5.90 on a Samsung Galaxy Z Flip8 and 1.95 on a Windows machine running on the CPU alone. A token is a small chunk of text, often part of a word. Peak memory was around 5.5 GB on the Mac and 7 GB on the phones in these tests. 169Pi does not state minimum hardware requirements, and a file size is not the same as the memory every setup will need.
In practice, around 14 tokens per second on a Mac is readable as the text streams in, while the phone figures suit short questions better than long reports. The CPU-only number is the honest warning: without a GPU or Apple Silicon, a thinking model that writes long reasoning before answering will feel slow.
What does the 262K context window mean?
The context window is how much text the model can consider at once. Alpie Seed 9B supports 262,144 tokens natively, enough in principle for very long documents. However, 169Pi’s own profiling only went up to 32,000 tokens, and the research page warns against extrapolating results beyond that. Memory also grows with context: in the BF16 tests, peak GPU memory rose from about 17.9 GB at 1,000 tokens to about 26.3 GB at 32,000 tokens.
My reading is that the tested range, not the headline maximum, is the number to plan around. If you want to feed it a whole report on a 16 GB laptop, split the work into sections and check the answers; treat the full 262K window as something to test on your own hardware before relying on it.
What local use changes, and where it stops
Running a model locally means your prompts and files are processed on your device rather than on a provider’s server. 169Pi describes this as allowing data to remain within an organisation’s own environment, which is a possibility created by the setup, not a formal guarantee: the apps, plugins and storage around the model still decide what leaves your machine.
There are other open questions. 169Pi has not published the composition of its training data, and its device numbers are not expected to be identical across hardware. For a team that handles sensitive documents, Seed is worth a careful trial precisely because the weights are open and the licence is permissive; the sensible test is your own documents, your own hardware and your own tasks, compared against the model you use today.
