Not tested by us yet. Alpie Seed shipped as open weights and GGUF files with no hosted API we can call, and we have not run the weights ourselves. Everything below comes from the model card on Hugging Face and a launch write-up on DEV. We'll update this post when we have run it against its base model on our own coding prompts.
Alpie Seed is a 9B-parameter, Apache 2.0 model from the Indian startup 169Pi, released on November 6, 2026 and aimed at coding and reasoning on a laptop or phone with no internet connection. The Hugging Face card says it is a fine-tune of Qwen3.5-9B, so the real question is how much it adds on top of its base. The headline benchmark numbers come from 169Pi itself and nobody has reproduced them yet.
| Alpie Seed | |
|---|---|
| Maker | 169Pi |
| Released | 2026-11-06 |
| Parameters | 9B |
| Base model | Qwen3.5-9B (per the model card) |
| Context | 262K native on the base model, only 32K tested on the card |
| Inputs | Text and images (vision needs a matching projector file and a recent llama.cpp) |
| License | Apache 2.0 |
| GGUF sizes | Q2_K 3.83 GB, Q3_K_M 4.62 GB, Q4_K_M 5.63 GB |
| Price | Free weights, you pay for your own hardware |
What is new versus Qwen3.5-9B
The model card describes capabilities in reasoning, math, coding and debugging, OCR, document understanding and chart reading. It does not describe the training data or method, and links to a separate primary card for that. The DEV launch article adds that the model was trained in 4-bit from the start and handles video, but neither claim appears on the GGUF page we read, so treat both as unconfirmed.
The one hard difference we can point to is packaging. 169Pi publishes ready-made GGUF files, an Ollama tag (ollama run 169pi/alpie-seed, with q2, q3 and vision variants, the vision one at 6.6 GB), and lists LM Studio, Jan and Docker Model Runner as supported runtimes. A phone demo runs in the PocketPal app.
The vendor-reported numbers
The launch article quotes these, and it says they are the vendor's own and not independently verified:
| Benchmark | Alpie Seed | Qwen3.5-9B | Sarvam 30B |
|---|---|---|---|
| SWE-bench Verified | 61.2% | 60% | 34% |
| LiveCodeBench v6 | 72.99% | 65.6% | 70% |
Read this table carefully. On SWE-bench Verified the gap over its own base model is 1.2 points, small enough to vanish with a different scaffold or sampling setup. On LiveCodeBench the lead over the base is larger, about 7 points. Sarvam 30B, a much bigger model, posts a lower SWE-bench score here, but only a 3-point deficit on LiveCodeBench, so "small beats big" holds for one benchmark and not the other, as the article itself points out.
The Hugging Face page adds perplexity figures: 9.0065 for Alpie Seed against 9.0251 for Qwen at BF16. That is a language-modeling measure, and the card itself warns that lower perplexity does not guarantee better results on a specific task. It also reports a Q4 perplexity of 8.8921, lower than the BF16 figure, and says this probably means the reference checkpoint differs from the one used for the Q4 file. It should not be read as Q4 beating full precision.
Hardware and speed
Reported speeds are modest. The launch article gives about 20 tokens per second on an RTX 3050 laptop GPU, about 14 tokens per second on a 16 GB MacBook M2 and about 9.6 tokens per second on an iPhone 17 Pro. The model card's own BF16 run shows 41.37 output tokens per second, with peak GPU memory of 26.252 GB at 32K context and latency rising from 0.464 s at 1K tokens to 3.738 s at 32K. For quantized use, the card measures a 4.13% perplexity increase going from BF16 to NF4 while cutting the footprint from 17.527 GB to 7.232 GB, with 88.40% top-1 token agreement. No official hardware minimums are given.
Access, and what is not open yet
The weights, GGUF files and Ollama tag are public, and there is a web playground. Alpie Local, a coding agent built on the model that plans work, edits code and runs tests, is invite-only with a waitlist, and the VS Code extension is only planned. The model alone is open. The agent that makes it useful for repository work is not yet.
Who should care
- Teams that cannot send code to a cloud API and want an Apache-licensed local assistant. Compare it with the picks in our open-source LLMs for on-prem list before settling.
- Anyone already running Qwen3.5-9B locally. The upgrade is only worth the download if it beats the base on your own code, which the public numbers cannot tell you.
- Privacy-sensitive apps that need an on-device option, see our privacy-first picks.
Skip it if you need frontier-level reasoning or large refactors. A 9B model will not match cloud models there, and 169Pi is a young company with no track record to lean on. The same week another small open-weight release landed, which we covered in Kanana-2-v-decide explained. We plan to run Alpie Seed against Qwen3.5-9B on real bug-fix and test-writing prompts once we can host the weights, and this post will be updated with the result.