Chat model
Inkling-Small high · chat · Jul 2026
high reasoning · 16 questions per request
- Accuracy, all 2,198 questions
- 80.0%78.3% to 81.6%, 95% interval; 1,758 right
- Answers kept at 99% confidence
- 34 of 2,1980 wrong; share right 89.8% to 100.0%, 95% interval
- Cost, $ per 1,000 questionsper 1M questions
- $1.137$1,137Basis and price under Cost
- Latency
- 97.7 smedian request, 16 questions per request
How we called it
- Kind
- Chat model
- Access
- OpenRouter chat API, DeepInfra's endpoint (fp8), fallbacks off, output cap 65,536 tokens, reasoning included; no response format (the endpoint has no structured outputs), so the prompt adds a paragraph stating the answer layout
- Paid as
- Provider service
- Model ID
thinkingmachines/inkling-small- Parameters
- Mixture of experts: 276B total (counted from the released files: 265.96B), about 12B active per step. Size class uses the total (276B).mixture of experts: language model + multi-token prediction layer + vision and audio encoders, counted; active per token as the card states it; card: 276B total, 12B active; source: counted from the checkpoint's safetensors, thinkingmachines/Inkling-Small@8cc5877, thinkingmachines.ai/model-card/inkling-small/, huggingface.co/thinkingmachines/Inkling-Small; read 9 November 2026
- Confidence from
- Written estimate
- Run
- 866, 8 November 2026
Settings
- Reasoning
- high
- Request format
- 16 questions per request
- Requests at once
- 4
- Temperature
- not recorded
Run record: harness seed, harness commit and prompt version
- Harness seed
- 0
- Harness commit
b0118bbae7ac- Prompt version
254597b8be16
Results
| Questions | Accuracy | 95% interval | Right |
|---|---|---|---|
| All questions | 80.0% | 78.3% to 81.6% | 1,758 of 2,198 |
| Reading | 87.2% | 84.4% to 89.5% | 572 of 656 |
| Grammar | 76.9% | 74.7% to 78.9% | 1,186 of 1,542 |
Comparison group: opqr. Compare with another setup.
Invalid answers: 0; read from broken replies: 16; empty replies asked again: 3 (3 came back with answers, 48 answers). Without the rule: 64 invalid. 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 868): Invalid answers: 0; read from broken replies: 80; empty replies asked again: 2 (2 came back with answers, 32 answers). Without the rule: 112 invalid. (how)
Accuracy by topic
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
- 483 of 547, first English run (867)
- Accuracy
- 88.3%
- Non-hallucination with “I don't know”
- 0.0%: 0 “I don't know” of 60 not right (run 869)
- Answered with “I don't know”
- 100.0%: 547 of 547
- Error rate with “I don't know”
- 11.0%: 60 of 547 answers
- Cost, $ per 1,000 questionsper 1M questions, first English run
- 0.425425at list prices
- Latency, median request, first English run
- 25.3 s; 4 requests at once
- English run
- 867
- Machine, English runs
- hosted API
- Date
- 8 November 2026
- Request format
- 16 questions per request
- Reasoning
- high
- Prompt version
5c3a4849752e- Translations, SHA-256
bf0d47cd385a93050c65906977e199ef7e45d5848dbbdef2bc8d124f4d5abf68
Confidence
- Answers kept at 99% or more
- 34, of which 0 wrong (share right: 89.8% to 100.0%, 95% interval); 2,164 skipped, the model less than 99% sure
- Calibration gap
- 7.9 points
- AUROC, all questions
- 0.78 0.76 to 0.80
- AUROC, reading
- 0.77 0.72 to 0.82
- AUROC, grammar
- 0.77 0.74 to 0.79
- Brier score
- 0.145
- Accuracy of its surest answers
- most confident fifth: 96.6% right; most confident half: 92.4% right; most confident four fifths: 86.9% right
- Typical score of a right answer
- 91.0%
- Typical score of a wrong answer
- 82.0%
Repeats
We did not repeat this setup.
Latency
- Latency, median request
- 97.7 s (this run)
- Latency, 90th percentile
- 173.4 s
- Requests at once
- 4
- Throughput
- 31.8 questions per minute of request time
- Request time
- 69.2 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
- $1.137$1,137
- Whole run
- $2.50
- Basis
- Estimated from published prices and recorded token usage
- Price
- $0.45 input, $0.10 cached input, $1.20 output per 1M tokens (source, 2026-11-08)
- Input tokens per question
- 248
- Output tokens per question
- 855