Meet Metis Hybrid: one model that decides, explains and chats

Most AI products run two models: a classifier that makes decisions and a chatbot that talks. Metis Hybrid, our new decision model, does both. It decides in milliseconds with an honest probability for every option, explains its decision when you ask, and chats like a normal assistant, all from the same weights.

One model instead of two

Teams that route tickets, check policies or score leads usually host a classifier for the decision and a separate chatbot for the conversation. That is two models to host, update and pay for.

Metis Hybrid is one model with two outputs. A small decision head picks an answer from a fixed set of options and gives a probability for each. The language side of the same model can explain that decision, answer in words or chat. Switching from deciding to chatting is a setting on the request, not a second server.

Mode	What you get	Time per request
Decide only	The answer plus a probability for every option	about 0.06 s
Decide + explain	The same, plus a short written explanation	about 3 to 5 s
Text only	An answer in words, no probabilities	about 3 to 5 s
Chat	A normal assistant conversation	depends on length

Times are from our live demo on one cloud GPU (NVIDIA L4).

Three kinds of question

Like Metis, Metis Hybrid answers questions with a fixed set of answers:

• Pick one option: which team gets this ticket?
• Yes or no: is this claim within policy?
• A score on a scale: how severe is this incident?

You write the question and the options, and it decides. There is no retraining per task.

Decisions in milliseconds

A decision is one pass over the text with no text written: about 60 milliseconds. Having the same model write its answer out takes about 3 seconds on the same GPU, so a decision uses about 50× less GPU time. That suits high volumes such as tickets, emails and reviews, and less compute per decision fits what GreenPT is for.

It tells you how sure it is

Every answer comes with a probability, and at the top of the scale you can trust it. When Metis Hybrid 4B was 90% sure or more on the JevBench questions, it was right 97% of the time (147 of 151). The 27B shows the same pattern, at 98%.

When Metis Hybrid 4B is 90%+ sure, it is right 97% of the time

147 of 151 answers given with 90%+ confidence were right, 231 JevBench questions

Below 90%, the model is more sure of itself than it should be: answers given at 70 to 85% were right about two times in three. So draw the line at 90%: send those answers straight through and route the rest to a person. And when a case is close, it says so. Scoring a sales lead, it rated “ready to buy” at 57% and “hot” at 42%: a close call, shown honestly.

Explains its decision on request

One switch adds a short explanation in plain language. Keep in mind what that explanation is: it is written after the decision and describes it. It is not proof that the decision is right, so the probability stays the number to act on.

Benchmarks

JevBench tests decision models on business questions. On its 231 public questions, none of which our models trained on, Metis Hybrid 4B scores 84% and the 27B scores 88%. Our own run on 14 October 2026 matched the published numbers within one question. The 27B’s lead comes from the hardest questions.

Every 4B version scores the same; the 27B leads on hard questions

Accuracy per JevBench difficulty tier, 231 public questions

Compressing the 4B costs almost nothing: the 4.5 GB version gives the same answer as the full model on 99.6% of the questions, and the 2.7 GB version on 94.4%, within one question of the full model’s score.

Runs on your own hardware

Version	Size	Runs on	JevBench
Metis Hybrid 4B	9.3 GB	A server or cloud GPU with 16 GB or more	84.4%
Metis Hybrid 4B Q8_0	4.5 GB	Laptops, Macs and edge servers, with llama.cpp	84.4%
Metis Hybrid 4B Q4_K_M	2.7 GB	The smallest footprint, with llama.cpp	84.0%
Metis Hybrid 27B	about 54 GB	A large GPU with 80 GB or more	88.3%

The compressed versions are small enough to run on a laptop, so the data never has to leave the device. All versions are built on open Qwen models and are in the Metis Hybrid collection on Hugging Face. The collection also holds a 9B and developer versions of the 4B; their model cards on Hugging Face have the full benchmark.

What Metis Hybrid does not do

Metis Hybrid is not right every time: 84% for the 4B and 88% for the 27B on JevBench, and lower on kinds of task it has never seen. It decides in one pass, so it is weak at calculations and dates; put computed figures in the case. It was trained mostly on English and also works on Dutch input.

Try Metis Hybrid

Find the models in the Metis Hybrid collection on Hugging Face. Want to know how Metis started? Read Meet Metis.