Instruction following

IFEval

Verifiable instructions — word counts, formats, forbidden words — checked by a program rather than a judge model.

IFEval is scored as a percentage of questions answered correctly. 15 of the 159 models in this index report a score for it. The highest published figure here is 92.9%, from Mistral Small 3.2 24B.

The benchmark itself All models

Reported scores

Top 15 on IFEval

Ordered by the figure each maker published. Models that have not reported this benchmark are not listed — an absent score is not a low score.

Models ranked by published IFEval score
#	Model	Published score
1	Mistral Small 3.2 24BMistral AI	92.9%Best
2	Llama 3.3 70BMeta AI	92.1%
3	Command ACohere	90.9%
4	Llama-3.1-Nemotron-Ultra-253BNVIDIA	89.5%
5	Llama 3.1 405BMeta AI	88.6%
6	Jamba 1.6 LargeAI21 Labs	88%
7	xLAM-2-70BSalesforce AI	88%
8	Llama 3.1 70BMeta AI	87.5%
9	GPT-4.1OpenAI	87.4%
10	Palmyra X5Writer	87%
11	Granite 3.3 8BIBM	85.9%
12	Llama 3.1 8BMeta AI	80.4%
13	Command R7BCohere	77.9%
14	Llama 3.2 3BMeta AI	77.4%
15	Llama 3.2 1BMeta AI	59.5%

Where these numbers come from

Every score on this page is a published figure, taken from the model's own card, system card, technical report or release post, or from a public leaderboard. CorX Labs did not run these evaluations. Most are self-reported by the lab that built the model, which means they were produced under that lab's own choice of prompt, scaffold and number of attempts — so treat them as a starting point for a shortlist, not as a settled ranking.

A score someone other than the model's maker measured is marked Independent and names its measurer. Those are the stronger numbers on this page — an outside harness has no reason to flatter anyone — and there are not many of them.

Where a figure has not been published, the cell reads Not reported rather than an estimate. Nothing here is inferred, interpolated or guessed. Each model records the month its row was last checked. Full method and caveats.

How to read this score

IFEval checks instructions a program can verify — write under 200 words, answer in JSON, never use the letter e. Because a script grades it rather than a judge model, the score is unusually reproducible.

Where it is weak

Following a format is not the same as following intent. A model can score highly here and still miss what was actually being asked for.