Chat model

LFM2.5-2.6B default reasoning · chat

default reasoning · 1 question per request

Accuracy, all 2,198 questions
33.9%31.9% to 35.9%, 95% interval; 745 right
Answers kept at 99% confidence
Not measureda plain answer line, no confidence
Cost, $ per 1,000 questionsper 1M questions
$2.761$2,761Basis and price under Cost
Latency
51.2 smedian request, 1 question per request

Back to the leaderboard.

How we called it

Kind
Chat model
Access
our GPU (logic's server, vLLM v0.30.0 BF16, revision LiquidAI/LFM2.5-2.6B@7141b56), one question per request, plain answer line CEVAP: X (no answer schema); sampling repetition_penalty 1.1, seed 0, temperature 0.1, top_k 50; output cap 32,768
Paid as
Our own hardware
Model ID
lfm2.5-2.6b
Parameters
2.69B dense, as the maker states it (counted from the released files: 2.70B).language model (tied embeddings, text only); source: counted from the checkpoint's safetensors, LiquidAI/LFM2.5-2.6B@7141b56, huggingface.co/LiquidAI/LFM2.5-2.6B; read 12 November 2026
Hardware
1 × RTX PRO 6000 Blackwell (96 GB); BF16
Confidence from
Not measureda plain answer line, no confidence
Run
1037, 12 November 2026

Settings

Reasoning
default
Request format
1 question per request
Requests at once
32
Temperature
not recorded
Run record: harness seed, harness commit and prompt version
Harness seed
0
Harness commit
6ad0eac51a4a
Prompt version
ba455bb00e5c

Results

QuestionsAccuracy95% intervalRight
All questions33.9%31.9% to 35.9%745 of 2,198
Reading45.4%41.7% to 49.3%298 of 656
Grammar29.0%26.8% to 31.3%447 of 1,542

Comparison group: ξπσ. Compare with another setup.

Invalid answers: 40; read from broken replies: 0; empty replies asked again: 0. Still invalid: 9 cut off at the output limit, 31 unreadable. An empty reply is asked once more; in a broken reply, any letter that is clearly readable is scored and the rest count as wrong. Invalid answers count as wrong (how).

With “Bilmiyorum” (run 1038): Invalid answers: 33; read from broken replies: 0; empty replies asked again: 0. Still invalid: 14 cut off at the output limit, 19 unreadable. (how)

Accuracy by topic
TopicPartAccuracyRight
Word and phrase meaningsReading33.6%44 of 131
Sentence meaningsReading49.4%76 of 154
How a writer explains and develops ideasReading50.0%35 of 70
Topic and main point of a paragraphReading60.2%59 of 98
How a paragraph fits togetherReading35.8%38 of 106
Details that support a paragraph’s main pointReading47.4%46 of 97
Word types, such as nouns and verbsGrammar32.1%45 of 140
Groups of words built around a nounGrammar18.5%12 of 65
Verbs and their formsGrammar36.3%33 of 91
Verb forms used like nouns or modifiersGrammar27.8%20 of 72
Word endings and what they doGrammar21.5%17 of 79
How words are builtGrammar27.5%19 of 69
The roles of words and phrases in a sentenceGrammar37.2%54 of 145
How verbs show who acts and who receives the actionGrammar28.4%21 of 74
Types of sentencesGrammar28.2%24 of 85
Questions combining several grammar topicsGrammar23.2%51 of 220
Sound patterns and changesGrammar26.0%20 of 77
SpellingGrammar21.8%29 of 133
PunctuationGrammar35.9%42 of 117
Problems in sentence wordingGrammar34.3%60 of 175

English

The Text tab's 547 kept reading questions translated into English, asked with this setup's Text tab settings. Every setup: English reading; against Turkish: English and Turkish.

Right answers
404 of 547, first English run (1034)
Accuracy
73.9%
Non-hallucination with “I don't know”
15.7%: 21 “I don't know” of 134 not right (run 1035)
Answered with “I don't know”
95.6%: 523 of 547
Error rate with “I don't know”
21.6%: 113 of 523 answers
Cost, $ per 1,000 questionsper 1M questions, first English run
0.964964at the maker's list price
Latency, median request, first English run
15.0 s; 32 requests at once
English run
1034
Machine, English runs
RTX PRO 6000
Date
12 November 2026
Request format
1 question per request
Reasoning
default
Prompt version
ad6f4cc346df
Translations, SHA-256
bf0d47cd385a93050c65906977e199ef7e45d5848dbbdef2bc8d124f4d5abf68

With “I don't know” (run 1035): Invalid answers: 3; read from broken replies: 0; empty replies asked again: 0. Still invalid: 3 unreadable. (how)

Confidence

Not measured: this setup was asked for a plain answer line, so it states no confidence and reports no option probabilities. It has no answers kept at a confidence threshold, no calibration gap, ranking score (AUROC) or Brier score, and no typical score of a right or wrong answer.

Repeats

We did not repeat this setup.

Latency

Latency, median request
51.2 s (this run)
Latency, 90th percentile
96.9 s
Requests at once
32
Throughput
33.2 questions per minute of request time
Request time
66.3 minutes, over the requests whose answers were kept

Latency: the time from sending one request to receiving its answer, measured from our computer, network and any retries within the request included. Request time: the time at least one request was in flight, over the requests whose answers were kept. It is not the run's elapsed time: it leaves out requests that failed and any pause before a run was resumed.

Cost

Cost per 1,000 questionsper 1M questions
$2.761$2,761
Whole run
$6.07
Basis
Estimated at Liquid AI's list price for the same model (FP8 at its endpoint); our hardware cost is not measured.
Price
$0.35 input, $0.50 output per 1M tokens (source, 2026-11-12)
Input tokens per question
273
Output tokens per question
5,331