Decision model
GLiDE 0.8B
1 question per request
- Accuracy, all 2,198 questions
- 29.3%27.4% to 31.2%, 95% interval; 643 right
- Answers kept at 99% confidence
- 18 of 2,1981 wrong; share right 74.2% to 99.0%, 95% interval
- Cost, $ per 1,000 questionsper 1M questions
- No list priceBasis and price under Cost
- Latency
- 0.022 smedian request, 1 question per request
How we called it
- Kind
- Decision model
- Access
- GLiDE server (Fastino's release code), self-hosted
- Paid as
- Our own hardware
- Model ID
glide-0.8b- Variant
- local
- Parameters
- 750M dense, as the maker states it (counted from the released files: 0.75B).language model (no LM head) + 255-way readout; source: counted from the checkpoint's safetensors, fastino/GLiDE-0.8B@9eacdee, huggingface.co/fastino/GLiDE-0.8B; read 10 November 2026
- Hardware
- 1 × DGX Spark, GB10 (128 GB unified memory); BF16
- Note
- Fastino's released GLiDE 0.8B (fastino/GLiDE-0.8B at 9eacdee), on Qwen3.5-0.8B, run on our hardware through the server it ships. Its checkpoint holds no vision encoder, so its server takes text only.
- Confidence from
- Option score
- Run
- 923, 10 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
3b14502cefed- Prompt version
254597b8be16
Results
| Questions | Accuracy | 95% interval | Right |
|---|---|---|---|
| All questions | 29.3% | 27.4% to 31.2% | 643 of 2,198 |
| Reading | 42.4% | 38.7% to 46.2% | 278 of 656 |
| Grammar | 23.7% | 21.6% to 25.9% | 365 of 1,542 |
Comparison group: σφψ. Compare with another setup.
Accuracy by topic
Vision (preliminary)
Preliminary: these numbers are not final and may change. This setup is condition A of the Vision tab, which asks the 763 questions with underlined or numbered words as text (A), as a page image (B), as both (C) and through OCR models (D); the share answered right, its 95% interval in the number's tooltip.
| Model | A · text onlyaccuracy, higher is better | B · image onlyaccuracy, higher is better | C · text plus imageaccuracy, higher is better | Best OCR routeaccuracy, higher is better |
|---|---|---|---|---|
| GLiDE 0.8B Open · 0.8B | 29.6% | No image input | No image input | 27.4%via Chandra OCR 2 |
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
- 332 of 547, first English run (925)
- Accuracy
- 60.7%
- Non-hallucination with “I don't know”
- 21.7%: 49 “I don't know” of 226 not right (run 927)
- Answered with “I don't know”
- 91.0%: 498 of 547
- Error rate with “I don't know”
- 35.5%: 177 of 498 answers
- Cost, $ per 1,000 questionsper 1M questions, first English run
- No list price
- Latency, median request, first English run
- 0.023 s; 1 request at once
- English run
- 925
- Machine, English runs
- DGX Spark
- Date
- 10 November 2026
- Request format
- 1 question per request
- Reasoning
- off
- Prompt version
47f5776e71db- Translations, SHA-256
bf0d47cd385a93050c65906977e199ef7e45d5848dbbdef2bc8d124f4d5abf68
Confidence
- Answers kept at 99% or more
- 18, of which 1 wrong (share right: 74.2% to 99.0%, 95% interval); 2,180 skipped, the model less than 99% sure
- Calibration gap
- 24.8 points
- AUROC, all questions
- 0.62 0.59 to 0.65
- AUROC, reading
- 0.69 0.64 to 0.73
- AUROC, grammar
- 0.54 0.51 to 0.58
- Brier score
- 0.268
- Accuracy of its surest answers
- most confident fifth: 48.7% right; most confident half: 34.8% right; most confident four fifths: 31.1% right
- Typical score of a right answer
- 57.9%
- Typical score of a wrong answer
- 47.4%
Repeats
We did not repeat this setup.
Latency
- Latency, median request
- 0.022 s (this run)
- Latency, 90th percentile
- 0.029 s
- Requests at once
- 1
- Throughput
- 2538.1 questions per minute of request time
- Request time
- 0.9 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
- No list price
- Basis
- No dollar estimate; see hardware and time
- Input tokens per question
- 359
- 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 | 1.1% | −0.14 points | 1%† |
| GPT-6 Astra high · chat | 1.3% | −0.16 points | 1%† |
| GPT-6.1 Sol high · chat | 1.3% | −0.16 points | 1%† |
| GPT-6 Astra low · chat | 1.3% | −0.16 points | 1%† |
| GPT-6 Astra low · chat · 1 per request | 1.3% | −0.16 points | 1%† |
| Gemini 3.8 Flash low · chat · 1 per request | 0.9% | −0.10 points | 1%† |
| GPT-6.1 Sol low · chat · 1 per request | 1.3% | −0.16 points | 1%† |
| GPT-6.1 Sol low · chat | 1.3% | −0.16 points | 1%† |
| Claude Opus 5.5 high · chat | 1.1% | −0.14 points | 1%† |
| GPT-6.1 Terra high · chat | 1.3% | −0.16 points | 1%† |
| Claude Opus 5.5 low · chat · 1 per request | 1.0% | −0.12 points | 1%† |
| Gemini 3.8 Flash low · chat | 1.1% | −0.14 points | 1%† |
| GPT-6.1 Terra low · chat | 1.3% | −0.16 points | 1%† |
| Claude Opus 5.5 low · chat | 1.3% | −0.16 points | 1%† |
| Claude Sonnet 5.5 high · chat · 1 per request | 1.3% | −0.15 points | 1%† |
| Claude Sonnet 5.5 high · chat | 1.2% | −0.15 points | 1%† |
† Leaves out the first model’s hardware cost, which we have not estimated.