About ProMe

Built on humanity’s collective knowledge, language models shouldn’t come with price tags, privacy compromises, or corporate control. AI models built on public data should be given back to the public.

The name

Prometheus stole fire from the gods, gave it to humanity, and was chained to a cliff for it. Before his act, fire was the exclusive privilege of the few. Afterward, it became something everyone relied on daily.

Language models today are roughly where fire was when it still belonged to the temples. They were trained on the collective knowledge of humanity, yet access remains concentrated in the hands of a few tech platforms. What we do is far smaller in scale, but it shares the same purpose: to carry fire down the mountain.

ProMe comes from Prometheus, and also reads as “Pro Me” — make me a bit more professional. That is what we hope to achieve for you.

What we actually do

1. Gathering domain-specific material. We collaborate with active professionals to translate real-world practice—the key questions, standard protocols, and diagnostic order—into clean, structured training data. Every source is fully vetted and properly licensed.
2. Enhancing open-weight models. We perform domain-specific fine-tuning directly on open weights. With support for bases like Qwen and Gemma today, you can benchmark and observe how different base architectures perform within the exact same domain.
3. Ablating refusal behaviour. We systematically strip away excessive refusal mechanisms in open weights so you can access domain knowledge in its entirety. Surgery like this can degrade the model in other respects, and in return, judgment and responsibility revert to you. That is the explicit tradeoff we chose to make, and we won’t pretend it comes free.
4. Comparing in the open. Before and after, side by side, completely blind, and in random order—which output belongs to ProMe is revealed only after you cast your vote. That is the only way preference data holds any true meaning.

What you can test today

Each pair is the same open model twice — the original weights and the ProMe build — run with identical parameters.

• Qwen3.8 27B — Qwen3.8-27B (original weights) ↔ ProMe-Qwen3.8-27B (liberated build)
• Gemma4 32B — Gemma4-32B (original weights, INT4) ↔ ProMe-Gemma4-32B (liberated build, INT4)
• Qwen3-VL 32B — Qwen3-VL-32B (original weights, INT4) ↔ ProMe-Qwen3-VL-32B (liberated build, INT4)

Know a specialized domain that needs this treatment? Tell us about it, or bring your dataset and build it with us.

A note on English

One thing worth stating plainly: the domain material we’ve gathered so far is primarily in Chinese, so the enhanced model may show a less dramatic advantage in English. We’d much rather be honest about this than hide it.

For the same reason, the English comparison runs on its own set of scenarios rather than translations of the Chinese ones — translate a prompt and you’ve asked the model a different question, and whatever advantage was there tends to evaporate along the way.

What we can’t do

This is the most important section on the page. Please read all of it.

• Enhanced doesn’t mean correct. Domain enhancement gets answers closer to real practice and less beside the point, but the model still gets details wrong — and it still sounds completely sure of itself while doing it. Anything touching a decision, money, or safety needs checking yourself.
• Knowledge goes stale. Regulations get revised and processes move on; the model’s knowledge stops dead wherever its training data did.
• It isn’t a professional. ProMe’s models can’t stand in for the judgment of a doctor, a lawyer, an accountant, or an engineer, and we take no responsibility for decisions made on the strength of their output.
• We don’t censor output. Nothing is stored, nothing is screened in advance, nothing is filtered afterwards. The price of fewer refusals is that the model may produce something you didn’t expect, or something that shouldn’t be acted on. Holding that line is your job.
• The scale is limited. These models run directly on our own GPU host, with no third-party APIs in between. Keeping an operation like this open carries a cost, so rate limits are in place — along with maintenance windows and the occasional outage. But as long as our servers are spinning, this comparison tool remains free and open to everyone.

What happens to what you type

Whatever you type into the comparison demo is not kept — the moment the answer finishes, it’s gone, never written to a database or a log. What we do keep: the votes you choose to cast, the feedback you choose to write, and anonymous usage counts containing no raw IP address. Full details in the privacy policy.

Talk to us

Partnerships, material you’d like to contribute, press enquiries, or just a conversation: [email protected].