Chantless
Introduction
The point of a powerful tool is to give a person more freedom. Freedom to follow an idea, to try something unreasonable, to make a thing exactly the way they believe it ought to be. The people who do remarkable work care about details most people never notice. They know when a line is wrong, when an explanation is too complicated, when one more feature would make the product worse. We call that taste. It takes years to develop, and we believe it should give you an extraordinary advantage.
AI makes that advantage more interesting. Two people can sit down with the same model and come away with very different work. Someone with experience can see the promising idea buried in a mediocre answer. They know which question to ask next, which mistake matters, and when to say, “We can do better than this.” That judgment lets them reach further into what the machine can do. As intelligence becomes more accessible, knowing what to do with it becomes more valuable.
That’s the idea behind Chantless. Take the work you want to do. Establish what doing it well means. Shape the intelligence around that standard, keeping what the work needs and removing the unessential. Chantless means casting a spell without the incantation: the power is there, and the ceremony falls away. We’re building for people who have something of their own to express—and want the freedom to carry it further. Your taste. Multiplied.
How are we approaching this?
Results
All videos show work generated by an agentic specialist extracted from Qwen 3.8 27B.
| Specialist | Model | Benchmark | Specialist score | Big model score | As % of big model |
|---|---|---|---|---|---|
| Function calling | Qwen3-8B | BFCL single-call | 670 (66.5%) | 667 (66.2%) | 100.4% |
| Function calling, smaller | Qwen3-8B | BFCL single-call | 719 (71.4%) | 667 (66.2%) | 107.8% |
| Terminal agent | Qwen3.8-27B | Terminal-Bench 2.1, pass@1 | 77.8% | 74.7% | 104.1% |
| Terminal agent, smaller | Qwen3.8-27B | Terminal-Bench 2.1, pass@1 | 74.4% | 74.7% | 99.6% |
| Math | Qwen3.8-27B | AIME 2025, avg@8 | 96.67% | 99.17% | 97.5% |
| Redaction | Qwen3.8-27B | RedactionBench | 0.725 | 0.707 | 102.5% |
Specialist results are with runtime memory <20% of the full precision weights.
All results are at mask extraction on inference level attribution and requiring no post-training. Code to be released soon.
The road ahead
The tools for building your own specialists and forking, improving your own and works from others would all be possible in next releases coming soon.
