Research · Metis Hybrid
One model that decides, explains and chats.
A decision in about 60 milliseconds, with an honest probability for every option.
Most AI products run a classifier to decide and a chatbot to talk. Metis Hybrid does both from the same weights: it decides, explains its decision when you ask, and chats like a normal assistant.
One model instead of two
Deciding or chatting is a setting on the request, not a second server.
- 60 msPer decisionOne pass over the text on one cloud GPU, nothing written
- 50×Less GPU timeThan having the same model write its answer out
- 97%Right when 90%+ sure147 of 151 answers from the 4B on public JevBench questions
Four modes
Pick what each request needs.
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 explains, answers in words or chats. Times are from our live demo on one cloud GPU.
Decide only
The answer plus a probability for every option.
Decide and explain
The same, plus a short written explanation.
Text only
An answer in words, without probabilities.
Chat
A normal assistant conversation.
It tells you how sure it is
90% sure or more? Act on it. Below that? Pass it to a person.
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%.
Below 90% it is more sure of itself than it should be: answers given at 70 to 85% were right about two times in three. So send the confident answers straight through and route the rest to a person. When a case is close, it says so: scoring a sales lead, it rated "ready to buy" at 57% and "hot" at 42%.
Benchmarks
Compressing the 4B costs almost nothing.
JevBench tests decision models on business questions. Our own run on 14 October 2026 matched the published numbers within one question.
- 84%Metis Hybrid 4BOn 231 public JevBench questions, none of them in training
- 88%Metis Hybrid 27BIts lead comes from the hardest questions
- 99.6%Same answer at 4.5 GBThe compressed 4B against the full model
Runs on your own hardware
Small enough for a laptop, so the data never leaves the device.
All versions are built on open Qwen models. The Metis Hybrid collection on Hugging Face also holds a 9B and developer versions of the 4B, with the full benchmark on each model card.
Metis Hybrid 4B
9.3 GB. A server or cloud GPU with 16 GB or more.
Metis Hybrid 4B Q8_0
4.5 GB. Laptops, Macs and edge servers, with llama.cpp.
Metis Hybrid 4B Q4_K_M
2.7 GB. The smallest footprint, with llama.cpp.
Metis Hybrid 27B
About 54 GB. A large GPU with 80 GB or more.
What Metis Hybrid does not do
Know its limits.
It 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 does not calculate reliably.
It decides in one pass, so it is weak at calculations and dates. Put computed figures in the case.
Its explanation is not proof.
The explanation is written after the decision and describes it. The probability stays the number to act on.
English first.
It was trained mostly on English and also works on Dutch input.
Metis Hybrid, in short
How is Metis Hybrid different from Metis?
Metis only decides: it returns a probability for every answer and writes no text. Metis Hybrid makes the same kind of decision and can also explain it, answer in words or chat, from the same model.
Read about Metis →How fast is a decision?
About 60 milliseconds on a single cloud GPU, because a decision is one pass over the text with nothing written. An explanation or a written answer takes about 3 to 5 seconds.
Can I trust the probabilities?
The high ones, yes. When Metis Hybrid 4B said it was 90% sure or more on the JevBench questions, it was right in 147 of 151 cases. Below 90% it is too sure of itself, so route those answers to a person.
Can I run it on my own hardware?
Yes. The 4B comes in compressed versions of 4.5 GB and 2.7 GB that run with llama.cpp on laptops, Macs and edge servers, so the data never has to leave the device.
Does it work in other languages?
It was trained mostly on English and also works on Dutch input.