Models | Convergent

Hardware

Trained on a range of hardware.

We don't stick to one machine. Depending on the model and the experiment, we train on whatever hardware fits the job best:

Newest

Ministral 3 14B with the Quasar personality by Convergent.

Newest · strongest

752M parameters, fine-tuned from Qwen3.5-0.8B. Our strongest model in Convergent Bench 3.0.

Built on Zynthos-1

Built on Zynthos-1 for math, logic, code and web search. Released on request.

Father model

246M parameters. The father of the Convergent family.

Small Language Model

101M parameters, trained on a single RTX 3060 Ti.

Lattice · Standard

Knowledge base for Lattice: 27.3M facts, 96.2% answer rate. No neural network, every answer explained.

Small Language Model · new

Tavik-1 trained further: 25.2% on internal arithmetic, 52 of 100 logic tasks.

Small Language Model

20.4M parameters, 41 MB. Trained from scratch, answers in fluent English sentences. More in the post.

Small Language Model

254,000 parameters, about 0.5 MB. An experiment in how small a chat model can be. More in the post.

These models are no longer developed.

Lattice model

37.1M facts, 95.5% answer rate.

Research system

An earlier research system from Convergent.

Every release and retirement, from the first systems to what comes next.

May 21, 2026
Mach 9 Master
An earlier research system from Convergent.
Jun 3, 2026
Ares 1.7
Lattice model with 37.1M facts and a 95.5% answer rate.
Sep 15, 2026
Mach 9 Master retired
An earlier research system from Convergent.
Oct 6, 2026
Ares 1.7 retired
Lattice model with 37.1M facts.
Oct 14, 2026
Lattice
Thinking without a neural network: answers from readable facts.
Oct 19, 2026
Pico 1
Our first language model: 254,000 parameters, about 0.5 MB.
Oct 20, 2026
Tavik-1
20.4M parameters, trained overnight from scratch on one graphics card.
Oct 20, 2026
Rimo
Run open models on your own PC.
Oct 21, 2026
Pico 1.2
461,000 parameters.
Oct 21, 2026
Cobalt-1
101M parameters, trained on one RTX 3060 Ti.
Oct 22, 2026
Pico 1.3 Coding Large
6.9M parameters, focused on code.
Oct 22, 2026
Convergent 3.0
Our benchmark suite: coding, knowledge work, QA, HCE and HTML.
Oct 22, 2026
Pico 1.3 Small
Only 11,000 parameters.
Oct 23, 2026
Zynthos-1
246M parameters, the father of the Convergent family, trained on one NVIDIA H100.
Oct 24, 2026
Andromeda-1
Our newest and strongest model, built on Zynthos-1.
Oct 24, 2026
Tavik-1.3
Tavik-1 trained further for math and logic.
Oct 24, 2026
Quasar 1.0
Fine-tune of Qwen3.5-0.8B, our strongest model in Convergent Bench 3.0.
Oct 24, 2026
Quasar 1.3
Ministral 3 14B by Mistral AI with the Quasar personality by Convergent.