Colibri on Samarthya

The model that never doubted

Thu, 05 Nov 2026 21:44:00 -0500

Evening experiment: could the local Qwen3-Coder 30B on my server write a post for this blog?

Fell: Claude Code pointed at Colibri sent a 19,545-token opening request and was still reading it after 84 minutes. Stopped it. Colibri kept reading anyway.
Got up: sent the same brief as one plain request to Colibri and to Ollama. Both answered in 10 to 14 minutes with fluent, well-structured drafts.
Learned: I had told them to mark anything they were unsure of. Between them: one mark, 22 errors, two references to pages that do not exist, and one draft that would have broken the site build.

Wrote it up: Can a local 30B model write my blog?, with the version I trust: What model weights are.

Can a local 30B model write my blog? I fact-checked it

Thu, 05 Nov 2026 21:41:00 -0500

An intern who never says “I’m not sure” needs more checking than one who asks too many questions. I asked a local model to write a blog post for me, told it to mark anything it was unsure of, and then checked every line.

Fluent, fast enough, and wrong in a dozen places.

After comparing Ollama and Colibri on speed, I wanted to know whether a local model is good enough for real work. So I picked a job I do every week: writing a post for this blog. The topic was one I had just covered myself, what model weights are, so I had a researched, sourced version to check against.

Two apples, one model

Thu, 05 Nov 2026 19:33:00 -0500

Spent the day comparing two ways of running a model locally: Ollama, which I know, and Colibri, which I installed last night. Same model on both, Qwen3-Coder 30B.

Fell: the 19 GB Colibri download stalled overnight, and the Ollama one stalled at the very end and started again from 2%. My desktop ran out of memory twice with both tools and a browser open.
Got up: moved the test to my Linux server. The model arrived in five minutes there, and the engine built in 25 seconds.
Learned: on the same machine the two were close, about 10 tokens a second each. On my busy desktop one Colibri flag, --cap 128, made it 4.5 times faster.

Wrote it up: Ollama or Colibri? A beginner’s comparison. Next: point a coding agent at Colibri and see whether it keeps up.

Colibri on my desk: a 30B model writes a network scanner, slowly

Thu, 05 Nov 2026 19:00:00 -0500

A new colleague who works slowly but knows a lot is still useful, as long as someone checks the work. I gave a 30-billion-parameter coding model, running entirely on my own desktop, one ordinary job and checked everything it handed back.

The whole experiment on one card.

Colibri is a small inference engine (the program that runs a model) built for one trick: a mixture-of-experts model only uses a few of its sub-networks, called experts, for each token, so Colibri keeps the busy experts in memory and reads the rest from disk. That lets a model that would not fit in RAM run on an ordinary machine. I wanted to know what that feels like on a real task, not a benchmark.

Ollama or Colibri? A beginner's comparison on two of my machines

Wed, 04 Nov 2026 21:21:17 -0500

A Fuji and a Granny Smith are both apples, and you would still not bake with the one you snack on. Ollama and Colibri both run AI models on your own machine, but they were bred for different jobs.

Both apples. Not the same apple.

I have used Ollama for a while, and even built a small demo on it. Colibri is new to me: I installed it this week. So this is a comparison by someone who knows one tool well and is a beginner at the other, which is probably where you are too.