cd /news/artificial-intelligence/my-1 · home topics artificial-intelligence article
[ARTICLE · art-95940] src=promptcube3.com ↗ pub= topic=artificial-intelligence verified=true sentiment=· neutral

My 1.

Developer ThorOdinson246 released nl2sh, a fine-tuned Qwen2.5-Coder-1.5B model that converts natural language to shell commands, achieving 0.620 on InterCode-ALFA, outperforming the untuned Qwen2.5-Coder-7B's 0.613 and running at 31.9 tok/s on a laptop CPU with only 941MB file size. The model, trained on 125k pairs and quantized to Q4_K_M, is available on Hugging Face and GitHub under Apache-2.0, but lacks safety checkers, so it may provide destructive commands if asked.

read2 min views1 publishedAug 13, 2026
My 1.
Image: Promptcube3 (auto-discovered)

tar

flags for the tenth time this month drove me to actually build a solution. I wanted something that could live locally on my machine without eating all my RAM, so I went with a fine-tuned Qwen2.5-Coder-1.5B. After training on 125k natural-language to command pairs and quantizing to Q4_K_M, the resulting file is only 941MB.The performance on my laptop (i5-11320H, 4 threads) is surprisingly snappy. I'm seeing about 31.9 tok/s with a median query time of 0.59s, using roughly 1.6GB of RAM via llama.cpp. It's a tiny footprint for something that effectively replaces a search engine for CLI syntax.

Performance and Benchmarks #

I ran this through InterCode-ALFA to see where it actually stands against the bigger players. The results were unexpected:

Fine-tuned 1.5B: 0.620Untuned Qwen2.5-Coder-7B: 0.613GPT-4o: 0.73

While it's not hitting frontier-level intelligence, it's basically matching a 7B model's output while using a fraction of the parameters and running entirely on a CPU. There is a 3B version available that scores even higher if you have the overhead for it.

Deployment and Risks #

This is a raw tool, meaning there are almost no static safety checkers. If you tell it to delete your root directory, it will gladly give you the exact command to do it. It's a "trust the user" implementation, which is why I'm treating it as a practical tutorial in efficiency rather than a foolproof product.

For anyone wanting to do a deep dive into the weights or the training pipeline, I've made everything available.

The weights are hosted here:

huggingface.co/ThorOdinson246/nl2sh-1.5b-Q4_K_M

And the implementation code is here:

github.com/ThorOdinson246/whatisit-nl2sh

The whole project is under Apache-2.0. I'm looking for people to poke holes in the training method or suggest better ways to handle the natural-language mapping. If you've got a specific AI workflow where you need instant CLI generation without the latency of a cloud API, this is a solid starting point.

Next Why is the DeepSeek-V4-Pro-0813 repo acting so strange on →

a practical ChatGPT prompt guide, with plenty of directly applicable cases.

── more in #artificial-intelligence 4 stories · sorted by recency
── more on @thorodinson246 3 stories trending now
sponsored brought to you by zahid.host 4,200+ EU-deployed projects
reading about agents? ship yours in a single git push.

Run your AI side-project on zahid.host

EU-based hosting, git-push deploys, automatic HTTPS, no cold starts. Free tier with a custom domain — perfect for shipping the agent you just read about.

$git push zahid main
Live at https://your-agent.zahid.host
Get free account → Pricing
from €0/mo · no card required
LIVE [news/my-1] indexed:0 read:2min 2026-08-13 ·