Decision model
REAX 1.3
1 question per request
- Accuracy, all 2,198 questions
- 77.9%76.2% to 79.6%, 95% interval; 1,713 right
- Answers kept at 99% confidence
- 240 of 2,1980 wrong; share right 98.4% to 100.0%, 95% interval
- Cost, $ per 1,000 questionsper 1M questions
- $0.185$185Basis and price under Cost
- Latency
- 0.250 smedian request, 1 question per request
How we called it
- Kind
- Decision model
- Access
- REAX Engine server (reax13 mode, the maker's release code), self-hosted, FP8
- Paid as
- Our own hardware
- Model ID
reax13- Parameters
- Two models run together, option probabilities averaged 50/50: 57.36B in all, all used on every question.two full models: part G (Gemma 4 31B line, with vision encoder) + part Q (Qwen3.8-27B line, with vision encoder and 255-way readout); source: counted from the checkpoint's safetensors, decisionmodels/REAX-1.3@fa4bf7e, huggingface.co/decisionmodels/REAX-1.3; read 9 November 2026
- Hardware
- 1 × RTX PRO 6000 Blackwell (96 GB); FP8 (block-128)
- Note
- REAX's released REAX 1.3 weights (decisionmodels/REAX-1.3 at fa4bf7e), a decision model of two parts that both score every question; each option's probability is the 50/50 mean of the two parts'. Part G blends CharlieBT/gx31-decider (Gemma line, from tattrongvu/Jeffan-31b-1710) with PelaAI/KnowLine-31B-Gen1 (Gemma-4-31B); part Q blends gurujustin/sn92-gen-p5 (Qwen3.5-27B line, through baselabs/mammoth-1) with vllm-sr/d3-ultra. The model card calls REAX a decentralised lab on Bittensor subnet 92. It ran on our hardware through the maker's engine (the reax13 mode of reaxlabs/reax-engine at ff8f595, its own server, serve.py), both parts in FP8 (block-128) on one RTX PRO 6000 Blackwell, text only, one request at a time. Its card lists English only.
- Confidence from
- Option score
- Run
- 827, 6 November 2026
Settings
- Reasoning
- off
- Request format
- 1 question per request
- Requests at once
- 1
- Temperature
- not recorded
Run record: harness seed, harness commit and prompt version
- Harness seed
- 0
- Harness commit
d34eca45758f- Prompt version
254597b8be16
Results
| Questions | Accuracy | 95% interval | Right |
|---|---|---|---|
| All questions | 77.9% | 76.2% to 79.6% | 1,713 of 2,198 |
| Reading | 92.4% | 90.1% to 94.2% | 606 of 656 |
| Grammar | 71.8% | 69.5% to 74.0% | 1,107 of 1,542 |
Comparison group: rt. Compare with another setup.
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
- 506 of 547, first English run (829)
- Accuracy
- 92.5%
- Non-hallucination with “I don't know”
- 6.7%: 3 “I don't know” of 45 not right (run 831)
- Answered with “I don't know”
- 99.5%: 544 of 547
- Error rate with “I don't know”
- 7.7%: 42 of 544 answers
- Cost, $ per 1,000 questionsper 1M questions, first English run
- 0.169169from GPU time
- Latency, median request, first English run
- 0.229 s; 1 request at once
- English run
- 829
- Machine, English runs
- RTX PRO 6000
- Date
- 6 November 2026
- Request format
- 1 question per request
- Reasoning
- off
- Prompt version
47f5776e71db- Translations, SHA-256
bf0d47cd385a93050c65906977e199ef7e45d5848dbbdef2bc8d124f4d5abf68
Confidence
- Answers kept at 99% or more
- 240, of which 0 wrong (share right: 98.4% to 100.0%, 95% interval); 1,958 skipped, the model less than 99% sure
- Calibration gap
- 4.1 points
- AUROC, all questions
- 0.87 0.85 to 0.88
- AUROC, reading
- 0.94 0.91 to 0.96
- AUROC, grammar
- 0.83 0.81 to 0.85
- Brier score
- 0.121
- Accuracy of its surest answers
- most confident fifth: 100.0% right; most confident half: 97.5% right; most confident four fifths: 87.1% right
- Typical score of a right answer
- 90.7%
- Typical score of a wrong answer
- 48.4%
Repeats
We did not repeat this setup.
Latency
- Latency, median request
- 0.250 s (this run)
- Latency, 90th percentile
- 0.361 s
- Requests at once
- 1
- Throughput
- 224.8 questions per minute of request time
- Request time
- 9.8 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
- $0.185$185
- Whole run
- $0.41
- 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
- 0
- Output tokens per question
- 0
Going first in front of an LLM
Replay of recorded answers, cutoff chosen on half the questions, averaged over the splits. Savings are estimates from API prices; the Method page explains how.
| LLM behind it | Share taken | Accuracy change | Saving |
|---|---|---|---|
| Gemini 3.8 Flash high · chat | 33.6% | −0.09 points | 26% |
| GPT-6 Astra high · chat | 36.4% | −0.14 points | 33% |
| GPT-6.1 Sol high · chat | 38.2% | −0.17 points | 27% |
| GPT-6 Astra low · chat | 37.9% | −0.18 points | 34% |
| GPT-6 Astra low · chat · 1 per request | 34.4% | −0.13 points | 34% |
| Gemini 3.8 Flash low · chat · 1 per request | 35.5% | −0.10 points | 16% |
| GPT-6.1 Sol low · chat · 1 per request | 36.5% | −0.14 points | 33% |
| GPT-6.1 Sol low · chat | 37.9% | −0.18 points | 22% |
| Claude Opus 5.5 high · chat | 41.9% | −0.22 points | 39% |
| GPT-6.1 Terra high · chat | 41.6% | −0.20 points | 6% |
| Claude Opus 5.5 low · chat · 1 per request | 41.2% | −0.24 points | 40% |
| Gemini 3.8 Flash low · chat | 41.4% | −0.20 points | -27% |
| GPT-6.1 Terra low · chat | 41.7% | −0.21 points | -2% |
| Claude Opus 5.5 low · chat | 50.5% | −0.20 points | 47% |
| Claude Sonnet 5.5 high · chat · 1 per request | 56.0% | −0.27 points | 54% |
| Claude Sonnet 5.5 high · chat | 76.8% | −0.05 points | 72% |