Introducing Distribution Fine-Tuning v1

One of the biggest problems with state-of-the-art LLMs is writing quality.

Despite rapid improvements in most other dimensions, expert human writers and readers can still easily detect—and despise—several telltale signs of LLM prose. So disliked is LLM output among serious readers and writers that AI-generated text is now widely derided as “slop.” At the time of writing, we believe the stylistic poverty of LLM text generation is perhaps the single biggest obstacle preventing LLMs from transforming expert knowledge work at scale.

In recent months, we have pioneered a novel approach to LLM post-training that addresses the problem of “slop” and points the way toward unprecedented gains in stylistic quality and control.

Our new approach is called Distribution Fine-Tuning (DFT), and today we are announcing Deft, a company and product built to bring DFT into practical use.

You can try it now, for free, in the console.

The Problem: SFT Is Not All You Need

Supervised fine-tuning (SFT) can teach a model to follow examples, but its outputs often drift away from the distributional properties of the training data. That is why frontier models overuse certain phrases (“delving into a rich tapestry”), sentence shapes (“it’s not just silly, it’s stupid”), and punctuation marks (especially em dashes). It is also why their writing tends toward blandness and incoherence. The model may satisfy next-token prediction, but no serious writer would use that many contrastive negations and em dashes in a single piece. Nor would they rely so heavily on generic, logical-sounding transitions instead of choosing sequences of thought suited to the subject at hand. These qualities of writing can be recognized—and taught to models—only at the level of distribution.

The Solution: Distribution Fine-Tuning

To apply this intuition, we began by constructing a set of distribution-level measures that could be applied both to training texts and to model outputs. Where outputs deviated from the training texts, we found that writing quality could be improved by closing the gap. Token-distribution distance, for example, tracks overused words and phrases. Maximum Mean Discrepancy compares embedding distributions, helping show when model outputs are too generic or conceptually thin. We also used a series of LLM judges.

Our early testing convinced us that great writing is not only a sequence of locally plausible tokens but also a distributional object. A finished essay has a texture, rhythm, density, and level of detail that all frontier models currently miss. This is what Distribution Fine-Tuning makes possible:

• Distribution-quality gains. DFT improved MMD by 49 percent and JMQ by 63 percent compared with the SFT baseline.
• Enhanced creativity and depth. When evaluated by judge models on specific writing dimensions, DFT outputs showed a 164 percent improvement in creativity and a 146 percent improvement in meaningful detail and depth.
• Clearer communication. We recorded a 28 percent increase in coherence and a 16 percent increase in clarity.
• Indistinguishability from human writing. We ran a sample of 100 model outputs through the Pangram AI detector, and they scored as 100 percent human-written.

For Deft, the product implications are clear. At the limit, we believe that serious and original writers will be able to turn rough ideas and drafts into publication-grade prose in a single click. The quality of those outputs will still depend on the knowledge, creativity, and spirit the writer brings to the Deft Editor. But everything in editing and publishing that is mechanical and algorithmic will eventually be done by LLMs. That is not yet possible with today’s frontier models, but we believe Distribution Fine-Tuning is how we get there. The Deft Editor is how writers can begin moving in that direction today.

Try Deft Today

Today, we are opening access to our first publicly available DFT model. This initial release is a smaller 14B-parameter model, trained entirely on a local server. Though still a proof of concept, it is already commercially valuable to a small group of early customers, and we are eager to hear from more users.

We are making this model available for free as we continue to scale our compute and improve the product. You will be asked to sign up after your fifth generation.

The Deft Editor v1 is focused on essay-like prose, and you remain responsible for your ideas and intentions. But now your outputs can more closely match the distribution of high-quality human writing, rather than merely sounding like it—vaguely—at the sentence level.

Try DFT v1 for free in the console.