Chat model

LFM3-1.7B reasoning off · chat

reasoning off · 1 question per request

Accuracy, all 2,198 questions
28.3%26.5% to 30.3%, 95% interval; 623 right
Answers kept at 99% confidence
Not measureda plain answer line, no confidence
Cost, $ per 1,000 questionsper 1M questions
$0.646$646Basis and price under Cost
Latency
4.90 smedian request, 1 question per request

Back to the leaderboard.

How we called it

Kind
Chat model
Access
our GPU (logic's server, vLLM nightly 0.34.0rc1.dev302 (commit be47189f; no released vLLM serves LFM3) BF16, revision LiquidAI/LFM3-1.7B@0dc4268), one question per request, plain answer line CEVAP: X (no answer schema); sampling min_p 0.0, repetition_penalty 1.1, seed 0, temperature 0.1, top_k 50, top_p 1.0, thinking off; output cap 32,768
Paid as
Our own hardware
Model ID
lfm3-1.7b-vllm-nothink
Parameters
1.73B dense, as the maker states it (counted from the released files: 1.73B).language model (tied embeddings, text only); source: counted from the checkpoint's safetensors, LiquidAI/LFM3-1.7B@0dc4268, huggingface.co/LiquidAI/LFM3-1.7B; read 9 November 2026
Hardware
1 × RTX PRO 6000 Blackwell (96 GB); BF16
Confidence from
Not measureda plain answer line, no confidence
Run
1227, 15 November 2026

Settings

Reasoning
off
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
7b06c24e4bf8
Prompt version
ba455bb00e5c

Results

QuestionsAccuracy95% intervalRight
All questions28.3%26.5% to 30.3%623 of 2,198
Reading37.2%33.6% to 41.0%244 of 656
Grammar24.6%22.5% to 26.8%379 of 1,542

Comparison group: φψΓ. Compare with another setup.

Invalid answers: 21; read from broken replies: 0; empty replies asked again: 0. Still invalid: 17 cut off at the output limit, 4 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 1234): Invalid answers: 47; read from broken replies: 0; empty replies asked again: 0. Still invalid: 21 cut off at the output limit, 26 unreadable. (how)

Accuracy by topic
TopicPartAccuracyRight
Word and phrase meaningsReading26.0%34 of 131
Sentence meaningsReading47.4%73 of 154
How a writer explains and develops ideasReading42.9%30 of 70
Topic and main point of a paragraphReading49.0%48 of 98
How a paragraph fits togetherReading34.9%37 of 106
Details that support a paragraph’s main pointReading22.7%22 of 97
Word types, such as nouns and verbsGrammar27.1%38 of 140
Groups of words built around a nounGrammar33.8%22 of 65
Verbs and their formsGrammar30.8%28 of 91
Verb forms used like nouns or modifiersGrammar29.2%21 of 72
Word endings and what they doGrammar12.7%10 of 79
How words are builtGrammar17.4%12 of 69
The roles of words and phrases in a sentenceGrammar20.7%30 of 145
How verbs show who acts and who receives the actionGrammar24.3%18 of 74
Types of sentencesGrammar22.4%19 of 85
Questions combining several grammar topicsGrammar24.5%54 of 220
Sound patterns and changesGrammar14.3%11 of 77
SpellingGrammar24.8%33 of 133
PunctuationGrammar29.1%34 of 117
Problems in sentence wordingGrammar28.0%49 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
399 of 547, first English run (1223); historical paired baseline, run 1223 with Turkish
Accuracy
71.8%, the mean of 3 seeds; 70.9%–72.9% min–max of 3 seeds
Seed runs
seed 0: run 1223, 399 of 547 right; seed 1: run 1302, 388 of 547 right; seed 2: run 1303, 391 of 547 right
Non-hallucination with “I don't know”
11.6%: 18 “I don't know” of 155 not right (run 1225)
Answered with “I don't know”
96.7%: 529 of 547
Error rate with “I don't know”
25.9%: 137 of 529 answers
Cost, $ per 1,000 questionsper 1M questions, first English run
0.355355at the maker's list price
Latency, median request, first English run
4.69 s; 32 requests at once
English run
1223
Machine, English runs
RTX PRO 6000
Date
15 November 2026
Request format
1 question per request
Reasoning
off
Prompt version
ad6f4cc346df
Translations, SHA-256
bf0d47cd385a93050c65906977e199ef7e45d5848dbbdef2bc8d124f4d5abf68

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
4.90 s (this run)
Latency, 90th percentile
43.6 s
Requests at once
32
Throughput
116.2 questions per minute of request time
Request time
18.9 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
$0.646$646
Whole run
$1.42
Basis
Estimated at Liquid AI's list price for the same model (FP8 at its endpoint); our hardware cost is not measured.
Price
$0.08 input, $0.20 output per 1M tokens (source, 2026-11-09)
Input tokens per question
274
Output tokens per question
3,120