Chat model

LFM2.5-8B-A1B default reasoning · chat

default reasoning · 1 question per request

Accuracy, all 2,198 questions
21.0%19.3% to 22.7%, 95% interval; 461 right
Answers kept at 99% confidence
Not measureda plain answer line, no confidence
Cost, $ per 1,000 questionsper 1M questions
$1.050$1,050Basis and price under Cost
Latency
39.2 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/LFM2.5-8B-A1B@dd83f5c), one question per request, plain answer line CEVAP: X (no answer schema); sampling repetition_penalty 1.05, seed 0, temperature 0.2, top_k 80; output cap 32,768
Paid as
Our own hardware
Model ID
lfm2.5-8b-a1b
Parameters
Mixture of experts: 8.3B total (counted from the released files: 8.47B), about 1.5B active per step. Size class uses the total (8.3B).mixture of experts: language model (tied embeddings, text only), counted; the card says 8.3B total, 1.5B active; source: counted from the checkpoint's safetensors, LiquidAI/LFM2.5-8B-A1B@dd83f5c, huggingface.co/LiquidAI/LFM2.5-8B-A1B; read 9 November 2026
Hardware
1 × RTX PRO 6000 Blackwell (96 GB); BF16
Confidence from
Not measureda plain answer line, no confidence
Run
1288, 15 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
88e48d7a1055
Prompt version
ba455bb00e5c

Results

QuestionsAccuracy95% intervalRight
All questions21.0%19.3% to 22.7%461 of 2,198
Reading25.6%22.4% to 29.1%168 of 656
Grammar19.0%17.1% to 21.0%293 of 1,542

Comparison group: ΘΛΞΠ. Compare with another setup.

Invalid answers: 346; read from broken replies: 0; empty replies asked again: 0. Still invalid: 6 cut off at the output limit, 340 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 1298): Invalid answers: 388; read from broken replies: 0; empty replies asked again: 0. Still invalid: 4 cut off at the output limit, 384 unreadable. (how)

Accuracy by topic
TopicPartAccuracyRight
Word and phrase meaningsReading22.9%30 of 131
Sentence meaningsReading27.9%43 of 154
How a writer explains and develops ideasReading22.9%16 of 70
Topic and main point of a paragraphReading34.7%34 of 98
How a paragraph fits togetherReading20.8%22 of 106
Details that support a paragraph’s main pointReading23.7%23 of 97
Word types, such as nouns and verbsGrammar10.7%15 of 140
Groups of words built around a nounGrammar23.1%15 of 65
Verbs and their formsGrammar13.2%12 of 91
Verb forms used like nouns or modifiersGrammar23.6%17 of 72
Word endings and what they doGrammar16.5%13 of 79
How words are builtGrammar15.9%11 of 69
The roles of words and phrases in a sentenceGrammar20.0%29 of 145
How verbs show who acts and who receives the actionGrammar16.2%12 of 74
Types of sentencesGrammar17.6%15 of 85
Questions combining several grammar topicsGrammar18.2%40 of 220
Sound patterns and changesGrammar18.2%14 of 77
SpellingGrammar18.8%25 of 133
PunctuationGrammar20.5%24 of 117
Problems in sentence wordingGrammar29.1%51 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
385 of 547, first English run (886)
Accuracy
70.4%
Non-hallucination with “I don't know”
26.4%: 51 “I don't know” of 193 not right (run 887)
Answered with “I don't know”
90.5%: 495 of 547
Error rate with “I don't know”
28.7%: 142 of 495 answers
Cost, $ per 1,000 questionsper 1M questions, first English run
0.517517from GPU time
Latency, median request, first English run
19.4 s; 32 requests at once
English run
886
Machine, English runs
RTX PRO 6000
Date
9 November 2026
Request format
1 question per request
Reasoning
default
Prompt version
ad6f4cc346df
Translations, SHA-256
bf0d47cd385a93050c65906977e199ef7e45d5848dbbdef2bc8d124f4d5abf68

With “I don't know” (run 887): Invalid answers: 1; read from broken replies: 0; empty replies asked again: 0. Still invalid: 1 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
39.2 s (this run)
Latency, 90th percentile
77.8 s
Requests at once
32
Throughput
39.5 questions per minute of request time
Request time
55.6 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
$1.050$1,050
Whole run
$2.31
Basis
Estimated from GPU time: the run's request time x the GPUs its server held x Runpod's on-demand price for an RTX PRO 6000, $2.49 per GPU-hour (read 5 November 2026). A cloud-equivalent estimate, not measured spending.
Price
$2.49 per GPU-hour, 1 GPU, for the run's request time (source, 2026-11-05)
Input tokens per question
272
Output tokens per question
2,537