Ordalin
Submit a tool

Reviewed tool profile

Alpie Seed

169Pi's 9B open-weight vision-language model for coding and reasoning, built to run offline.

Sources checked

Visit website ↗
Pricing
The model is free: the FP16 weights and the 2-, 3- and 4-bit GGUF builds are Apache 2.0 and can be used commercially. 169Pi runs a hosted API with 5M free tokens per key, but Alpie Seed is still listed there as coming soon, and API pricing will be published when it opens. Official source 1 ↗ · Official source 2 ↗ · Official source 3 ↗
Available as
Download from Hugging Face (FP16 safetensors) and run with Transformers 5 or later, or serve it as an OpenAI-compatible API with vLLM. · GGUF builds for llama.cpp, plus Ollama tags from 3.8 GB (q2) to 6.6 GB (vision), and a Kaggle Models listing. · 169Pi lists iPhone, Android, Mac, Windows and NVIDIA GPUs, through llama.cpp, MLX, CoreML and PocketPal. · An online Alpie Playground for trying it without a local setup. Official source 1 ↗ Official source 1 ↗ Official source 1 ↗ Official source 1 ↗
Primary group
Coding & Development

Also consider

See all 6 similar tools ↓

What it does

Alpie Seed is a 9B vision-language model from 169Pi AI, an Indian edge-AI lab (Buddhimatrix Pvt. Ltd.), released in late October 2026. It is Qwen3.5-9B post-trained with supervised fine-tuning and GRPO reinforcement learning for reasoning, coding and visual understanding, with thinking and non-thinking modes and a 262K native context. 169Pi's pitch is local use: the 4-bit build is 5.24 GB and runs offline on laptops and phones. Weights are on Hugging Face under Apache 2.0, with GGUF builds and Ollama tags.

Official source 1 ↗ · Official source 2 ↗ · Official source 3 ↗

Features

  • Runs offline on a laptop or phone. The 4-bit build is 5.24 GB. 169Pi measured about 14 tokens per second on a 16 GB M2 Mac and about 9.6 on an iPhone 17 Pro, with no network connection. Official source 1 ↗ · Official source 2 ↗
  • Reasoning with thinking and non-thinking modes. Hybrid reasoning: switch extended thinking on for hard problems and off for quick answers. Official source 1 ↗
  • Coding. Writes, fixes and explains code offline. 169Pi reports 72.99 on LiveCodeBench v6 and 61.20 on SWE-Bench Verified, ahead of the Qwen3.5-9B base at 60.0. Official source 1 ↗ · Official source 2 ↗
  • Images, documents and video. Reads photos, screenshots, charts, scanned documents and video in the same model; 169Pi reports 90.30 on OCRBench. Official source 1 ↗ · Official source 2 ↗
  • Indian languages. 169Pi evaluated it in ten Indian languages, including Hindi, Bengali, Tamil and Telugu, reporting 73.67% on MMLU Indic overall. Official source 1 ↗
  • Long context and tool calls. A 262K-token native context, and it emits tool-call requests that your application executes. Official source 1 ↗

Pricing & free limits

Open weights

FreeDownload; Apache 2.0

FP16 weights and 2-, 3- and 4-bit GGUF builds on Hugging Face, Ollama and Kaggle; personal and commercial use allowed under the licence. Official source 1 ↗ · Official source 2 ↗

Enterprise deployment

Contact salesCustom

Private on-prem, VPC or air-gapped deployment with custom models and SLAs. Official source 1 ↗

Billing details

169Pi API keys can be recharged in rupees or dollars once paid usage opens for Alpie Seed. Official source 1 ↗

Free access & limits

The open weights have no usage limit; the hosted API starts every key with 5M free tokens once Alpie Seed is available there. Official source 1 ↗ · Official source 2 ↗

Best for & limitations

Coding help on a laptop with no internet connection or API key. Official source 1 ↗

Reading invoices, scans and screenshots on the device, without sending them to a server. Official source 1 ↗

Apps for Indian-language users that need to work on mid-range hardware. Official source 1 ↗

Private or air-gapped deployments where data cannot leave the building. Official source 1 ↗ · Official source 2 ↗

  • A fine-tune of Qwen3.5-9B. The architecture is unchanged from Qwen3.5-9B; 169Pi's work is post-training with SFT and GRPO adapters. Official source 1 ↗ · Official source 2 ↗
  • Vendor-reported, mixed-harness scores. Benchmarks are 169Pi's own, run on the FP16 model with extended thinking; comparison scores come from other model cards and were not re-run. Official source 1 ↗ · Official source 2 ↗
  • Slow on phones and CPU-only PCs. 169Pi's own tests show about 9.6 tokens per second on an iPhone 17 Pro and under 2 on a CPU-only Windows PC, with about 6–7 GB of memory in use. Official source 1 ↗
  • Quantized builds score lower than reported. Published scores do not apply to the 2-, 3- or 4-bit builds most people run locally. Official source 1 ↗ · Official source 2 ↗
  • Vision and long context depend on your runtime. Image input needs a compatible runtime and the vision components; the default Ollama tags are text-only, and the usable context depends on runtime and memory. Official source 1 ↗ · Official source 2 ↗
  • No hosted API yet. Alpie Seed is not yet served on 169Pi's API, so you have to run it yourself or use the playground. Official source 1 ↗ · Official source 2 ↗
Alpie Seed product screenshot
Website previewFixed desktop viewport · 1440 × 900

Similar tools

Published tools that share its categories.

OpenLLMA routing layer that puts your AI subscriptions, API keys and local models behind one key for Claude Code, Codex and other agents, with fallback chains.
Coding & DevelopmentFree planChecked Oct 31
Visit
ColdteaControl layer for AI-written software: map every pull request, test the flows it touches with agents and watch it in production.
Coding & DevelopmentFree planChecked Aug 23
Visit