Darren Labs Agentic AI Studio

Darren Labs AI Lab · First models · 24 October 2026

Our first
models.

Ran on our own VM in Batam, timed in public, slow results included.

7B answered in 16 seconds on CPU. 32B took about seven minutes on CPU. Both are free to run with Ollama.

Run them free See the models

First flagship
DarrenLabs AI Models
Models
7B and 32B, live on ai-lab
Licence
Apache 2.0 bases, open recipe
Speed shown
32GB VRAM average, CPU measured below

The models

Two models, same CPU machine, both timings shown.

7B

DarrenLabs-7B · Live

The small one, for notes and first drafts.

Turn a rough site note into a readable checklist, or a messy drawing note into a first draft you can edit. Based on Qwen2.5-7B-Instruct.

Size
4.7 GB
Speed, 32GB VRAM avg
169 tokens / sec
On our CPU VM
8.8 tokens / sec
Tokens
54
“I am DarrenLabs-7B, an assistant running on the DarrenLabs AI Lab VM. I am based on Qwen2.5-7B, an open-source model from Alibaba Qwen, used under the Apache 2.0 license.”

Use it for a first draft, then review it. On 32GB VRAM, the 54-token test answer generates in about 0.3 seconds on average, about 19 times faster than our CPU. On CPU, the test took 16.0 seconds in total.

32B

DarrenLabs-32B · Live, slow on CPU

The bigger one, for longer documents.

For a full drawing quality note, a checklist with reasons, or a second pass over a long document. Based on Qwen3-32B. Fast on a 32GB GPU, slow on CPU, and the slow time is on this page.

Size
20 GB
Speed, 32GB VRAM avg
54 tokens / sec
On our CPU VM
1.89 tokens / sec
Tokens
292, incl. thinking
“I am DarrenLabs-32B, an AI assistant deployed on the DarrenLabs AI Lab VM (Oracle Cloud). I am based on Qwen3-32B, an open-source model from Alibaba Qwen, licensed under Apache 2.0.”

Use it for longer work, then review it. On 32GB VRAM, the 292-token test answer generates in about 5.4 seconds on average, about 29 times faster than our CPU. On CPU, the same answer took 413.9 seconds, about seven minutes. CPU is measured. GPU is an average from published benchmarks, not yet our own run.

Speed

On 32GB VRAM.

7B: 169 tokens per second on average. 32B: 54. On CPU, measured: 8.8 and 1.89.

Average: 7B (150 + 185.9 + 169.8) / 3. 32B (45.1 + 61.4 + 55.5) / 3. Published RTX 5090 benchmarks. Our GPU run is still to come.

How it runs

How to run them.

No website chat. You pull the base model, create the DarrenLabs version, and run it on your own machine. That is why it is free, and why the 32B version is slow on a CPU-only machine.

You, at home ollama pull ollama create DarrenLabs AI Lab VM Oracle Cloud, Batam DarrenLabs-7B 169 tok/s avg, 32GB VRAM DarrenLabs-32B 54 tok/s avg, 32GB VRAM An answer checked by a human before it goes anywhere

Speeds: 32GB VRAM averages. CPU: 8.8 and 1.89, measured.

Run them free, anywhere

Run them.

Ollama on your own machine. No account, no fee, no data sent to us.

# 7B
ollama pull qwen2.5:7b
ollama create darrenlabs-7b -f Modelfile.darrenlabs-7b

# 32B, give it time on CPU, use the Ollama API for first tests
ollama pull qwen3:32b
ollama create darrenlabs-32b -f Modelfile.darrenlabs-32b

Free means the method and the model files, not a public chat on our machine. Our VM is not open to the internet for chat.

Use, and limits

Use them. Check them.

Use them for

  • Turning a rough note about a duct joint, a valve, or a cable tray into a clear checklist with items to check
  • A first draft of a drawing quality note, for an engineer to correct and approve
  • A first draft for your review, correction, and approval

Do not use them for, unchecked

  • Putting an unchecked number on a drawing
  • Sending an unchecked answer to a client
  • Any decision a licensed engineer has not reviewed

What is next

Next.

  1. Smaller student model. Trained on 50 teacher examples we generated on our own VM, then scored against a held-out set.
  2. Faster 32B. Thinking off first, then a 30B MoE model, or a stronger machine, whichever the timing proves fastest.
  3. More models. Only after the same public test: identity answer, timing, and limits, on this page.

Browse all projects Back home