Decision model
GLiDE 27B
1 question per request
- Accuracy, all 2,198 questions
- 71.1%69.2% to 73.0%, 95% interval; 1,563 right
- Answers kept at 99% confidence
- 247 of 2,1980 wrong; share right 98.5% to 100.0%, 95% interval
- Cost, $ per 1,000 questionsper 1M questions
- No list priceBasis and price under Cost
- Latency
- 0.542 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-27b- Variant
- local
- Parameters
- 27B dense, as the maker states it (counted from the released files: 26.09B).language model (no LM head) + 0.46B vision encoder + 255-way readout; source: counted from the checkpoint's safetensors, fastino/GLiDE-27B@4d32743, huggingface.co/fastino/GLiDE-27B; read 10 November 2026
- Hardware
- 1 × DGX Spark, GB10 (128 GB unified memory); BF16
- Note
- Fastino's released GLiDE 27B (fastino/GLiDE-27B at 4d32743, 5 November 2026), a modified Qwen3.8-27B whose layers 24 to 47 run twice (recurrent depth, its decision_config.json), run on our hardware through the server it ships. The same repository's earlier checkpoint, kept at its tag glide-v1.2, is the GLiDE · v1.2 row.
- Confidence from
- Option score
- Run
- 899, 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
ce3fd5527feb- Prompt version
254597b8be16
Results
| Questions | Accuracy | 95% interval | Right |
|---|---|---|---|
| All questions | 71.1% | 69.2% to 73.0% | 1,563 of 2,198 |
| Reading | 90.9% | 88.4% to 92.8% | 596 of 656 |
| Grammar | 62.7% | 60.3% to 65.1% | 967 of 1,542 |
Comparison group: yzAC. 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 27B Open · 27B | 69.1% | 49.9% | 68.8% | 64.2%via Surya 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
- 497 of 547, first English run (901)
- Accuracy
- 90.9%
- Non-hallucination with “I don't know”
- 37.9%: 22 “I don't know” of 58 not right (run 903)
- Answered with “I don't know”
- 96.0%: 525 of 547
- Error rate with “I don't know”
- 6.9%: 36 of 525 answers
- Cost, $ per 1,000 questionsper 1M questions, first English run
- No list price
- Latency, median request, first English run
- 0.545 s; 1 request at once
- English run
- 901
- 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
- 247, of which 0 wrong (share right: 98.5% to 100.0%, 95% interval); 1,951 skipped, the model less than 99% sure
- Calibration gap
- 2.3 points
- AUROC, all questions
- 0.84 0.83 to 0.86
- AUROC, reading
- 0.93 0.90 to 0.95
- AUROC, grammar
- 0.78 0.76 to 0.81
- Brier score
- 0.144
- Accuracy of its surest answers
- most confident fifth: 100.0% right; most confident half: 93.0% right; most confident four fifths: 80.1% right
- Typical score of a right answer
- 88.2%
- Typical score of a wrong answer
- 49.0%
Repeats
We did not repeat this setup.
Latency
- Latency, median request
- 0.542 s (this run)
- Latency, 90th percentile
- 0.676 s
- Requests at once
- 1
- Throughput
- 106.1 questions per minute of request time
- Request time
- 20.7 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 | 27.1% | −0.06 points | 27%† |
| GPT-6 Astra high · chat | 28.5% | −0.10 points | 29%† |
| GPT-6.1 Sol high · chat | 29.6% | −0.10 points | 30%† |
| GPT-6 Astra low · chat | 28.5% | −0.10 points | 29%† |
| GPT-6 Astra low · chat · 1 per request | 27.4% | −0.08 points | 27%† |
| Gemini 3.8 Flash low · chat · 1 per request | 29.9% | −0.07 points | 30%† |
| GPT-6.1 Sol low · chat · 1 per request | 32.4% | −0.13 points | 32%† |
| GPT-6.1 Sol low · chat | 28.5% | −0.10 points | 29%† |
| Claude Opus 5.5 high · chat | 30.7% | −0.19 points | 31%† |
| GPT-6.1 Terra high · chat | 30.2% | −0.13 points | 30%† |
| Claude Opus 5.5 low · chat · 1 per request | 29.5% | −0.10 points | 30%† |
| Gemini 3.8 Flash low · chat | 27.8% | −0.12 points | 28%† |
| GPT-6.1 Terra low · chat | 35.4% | −0.10 points | 35%† |
| Claude Opus 5.5 low · chat | 35.5% | −0.14 points | 36%† |
| Claude Sonnet 5.5 high · chat · 1 per request | 34.5% | −0.10 points | 35%† |
| Claude Sonnet 5.5 high · chat | 56.0% | −0.14 points | 56%† |
† Leaves out the first model’s hardware cost, which we have not estimated.