Recommendation for Vision / Documents
Vision & Documents
Our top recommendation for Vision & Documents, based on the public evidence we track, is Anthropic: Claude Opus 4.6.[1][2] Use for document parsing workflows where benchmark-proven quality matters, as it scores 1507 on LMArena's document arena. OpenAI: GPT-5.6 Sol is the next-ranked alternative. Deploy for sheet music transcription and other fine-grained visual pattern recognition tasks, as it accurately follows horizontal lines and staves that trip up other vision models.
About this recommendation
- Updated
- Sep 25, 2026
- Evidence through
- Sep 25, 2026
- Sources
- 14
- Revision
- v78
Decision audit
Why this result
Inspect the inputs and the computed order behind the recommendation.
Models screened
20
live candidates
Evaluation feeds
5
task-weighted
Winner coverage
38%
intended feed weight
Largest provider share
2 of 7
Anthropic
Sources evaluated
The task sets these weights before any model is scored.
| Evaluation feed | Weight | Winner result | Field measured |
|---|---|---|---|
| VLMEvalKit tasksunavailable | 40% | feed unavailable | 0/20 |
| LMArena Document | 20% | #3 | 12/20 |
| LMArena Vision | 15% | #3 | 16/20 |
| Structured-output evalunavailable | 15% | feed unavailable | 0/20 |
| OpenRouter usage | 10% | 83/100 | 20/20 |
Provider concentration
Each exact model is scored separately; provider identity is not a ranking input.
- Anthropic2 models
- deepseek2 models
- Google1 model
- OpenAI1 model
- Qwen1 model
Decision table
Every published model is shown in computed order. Practitioner sources are distinct community threads, not the citations repeated in the prose below.
| Rank | Model | Relative score | Coverage | Practitioner evidence | Strongest measured reason |
|---|---|---|---|---|---|
| 01 | Claude Opus 4.6Anthropic | 54 | 38% | 1 threads · 1 families · 0 cautions | #3 LMArena Document · #3 LMArena Vision |
| 02 | GPT-5.6 SolOpenAI | 52 | 38% | 1 threads · 1 families · 1 cautions | #6 LMArena Document · #8 LMArena Vision |
| 03 | Claude Sonnet 4.6Anthropic | 49 | 38% | 1 threads · 1 families · 1 cautions | #7 LMArena Document · #16 LMArena Vision |
| 04 | DeepSeek V4 Flash 0423deepseek | 49 | 11% | 6 threads · 6 families · 1 cautions | OpenRouter usage 98/100 normalized |
| 05 | DeepSeek V4 Flash Vision Expdeepseek | 47 | 11% | 15 threads · 12 families · 5 cautions | OpenRouter usage 86/100 normalized |
| 06 | Qwen3.8 27BQwen | 47 | 28% | 4 threads · 4 families · 1 cautions | #31 LMArena Vision |
| 07 | Gemini 3.5 Flash LiteGoogle | 47 | 28% | 2 threads · 2 families · 1 cautions | #20 LMArena Vision |
Relative score combines normalized benchmark quality and signal coverage; independent practitioner evidence and freshness are bounded tie-breakers. It is an ordering score, not an absolute quality percentage. The writing model receives this order and cannot change it.
Claude Opus 4.6 ranks third on LMArena's document arena with a score of 1507, indicating strong blind preference performance on document tasks.
Best when: Use for document parsing workflows where benchmark-proven quality matters, as it scores 1507 on LMArena's document arena.
Tips
- Use for document parsing workflows where benchmark-proven quality matters, as it scores 1507 on LMArena's document arena.
GPT-5.6 Sol ranks ninth on LMArena's vision arena with Elo 1286, and community evidence shows it successfully transcribes sheet music where other models fail.
Best when: Deploy for sheet music transcription and other fine-grained visual pattern recognition tasks, as it accurately follows horizontal lines and staves that trip up other vision models.
Tips
- Deploy for sheet music transcription and other fine-grained visual pattern recognition tasks, as it accurately follows horizontal lines and staves that trip up other vision models.
Claude Sonnet 4.6 ranks eighth on LMArena's document arena with score 1482, and was evaluated in a 26,904-query food photography study at temperature 0.01.
Best when: Consider for structured visual extraction tasks requiring consistent outputs, as it was tested with temperature 0.01 in a large-scale food photography analysis pipeline.
Tips
- Consider for structured visual extraction tasks requiring consistent outputs, as it was tested with temperature 0.01 in a large-scale food photography analysis pipeline.
DeepSeek V4 Flash 0423 is a text-only model with no native vision capability; the vision variant is a separate experimental model, and evidence shows numeric extraction errors on OCR-digitized documents.
Best when: Consider only after reviewing the cited caution.
Watch out for
- Do not use for any image or document tasks; the hosted API is text-only and lacks vision support entirely.
- Avoid even for OCR-then-process workflows, as ironclaw logs show 38 genuine model-quality failures on numeric extraction from OCR-digitized Treasury Bulletins.
DeepSeek V4 Flash Vision Exp is an experimental multimodal model released 2026-08-21 with OpenAI-compatible vision API, supporting base64 data URLs and external image URLs up to 32 MiB.
Best when: Use for cost-sensitive vision workloads, as production logs show 9–13 second latency per successful visual qualification call.
Tips
- Use for cost-sensitive vision workloads, as production logs show 9–13 second latency per successful visual qualification call.
- Deploy for long-context vision tasks, as independent validation confirms 200K-token context window with 9/9 needle-in-haystack passes.
Watch out for
- Watch for model name normalization issues in tooling; some clients strip the -vision-exp suffix and route to the text-only variant.
- Note the experimental status and routing changes; V4 Flash Vision Exp has been deprecated with names temporarily routing to V4.1 Flash.
Qwen3.8 27B is a 27B dense hybrid-attention model with vision tower, ranking 27th on LMArena's vision arena with Elo 1251, and achieves 122 tok/s on single H200 GPU at full 262K context.
Best when: Use for high-throughput local vision deployment, as it fits one H200 GPU at full 262144 context with 122 tok/s single-caller throughput and scales to 3957 tok/s at 1024 concurrency.
Tips
- Use for high-throughput local vision deployment, as it fits one H200 GPU at full 262144 context with 122 tok/s single-caller throughput and scales to 3957 tok/s at 1024 concurrency.
- Consider for memory-constrained edge deployment, as Unsloth GGUF quantization brings the 27B dense model to 5.95 GB at 1.75 bpw plus 0.63 GB vision projection.
Watch out for
- Verify provider configuration carefully; production logs show ID extraction failures when promoted to default on Groq due to vision capability mismatches.
Gemini 3.5 Flash Lite ranks 23rd on LMArena's vision arena with Elo 1266, and is used as a fallback when quota limits block newer Gemini versions.
Best when: Configure as automatic fallback for free-tier receipt scanning workflows when primary models hit quota, as it maintains service continuity without user-visible delays.
Tips
- Configure as automatic fallback for free-tier receipt scanning workflows when primary models hit quota, as it maintains service continuity without user-visible delays.
Watch out for
- Switch to image mode instead of whole_pdf OCR when processing documents, as Gemini frequently rejects PDF OCR requests that succeed in image mode.
- Handle empty response retries in your pipeline, as logs show 'candidate with no content parts' errors that exhaust configured retry limits before falling through to backup providers.
Frequently asked
- What is the top-ranked model for Vision & Documents?
- Anthropic: Claude Opus 4.6 ranks first in the current evidence-weighted comparison. Use for document parsing workflows where benchmark-proven quality matters, as it scores 1507 on LMArena's document arena.[1]
- What is an alternative to Anthropic: Claude Opus 4.6?
- OpenAI: GPT-5.6 Sol is the next-ranked option. Deploy for sheet music transcription and other fine-grained visual pattern recognition tasks, as it accurately follows horizontal lines and staves that trip up other vision models.[2]
Sources
- 1
“Ranks #3 of 34 on LMArena's document arena (score 1507), based on blind preference for document tasks.”
LMArena document arena · Benchmark · Sep 13, 2026 - 2
“So far I haven't seen a single model succeeding at transcribing sheet music, but I just tested it again with 5.6 Sol and it nailed the small test case. Fluently reading music requires multiple years of training for most people, but I feel like accurately following the horizontal lines trips up vision models in particular.”
kherud · Hacker News · Aug 17, 2026 - 3
“The study used a temperature of 0.01. > "Thirteen food photographs were each submitted 495–561 times to four LLM vision APIs (GPT-5.4, Claude Sonnet 4.6, Gemini 2.5 Pro, Gemini 3.1 Pro Preview) using an identical structured prompt adapted from the iAPS automated insulin delivery system (26,904 total queries, temperature 0.01)"”
muwtyhg · Hacker News · Apr 29, 2026 - 4
“## Standing order Do **not** OCR-then-DeepSeek / text-only 2.4T. DeepSeek v4-flash was unscored because the hosted API is text-only. If a VL host for Qwen3.8-2.4T-A95B appears, run the same seed-777 JSONL + scorer as Luna/restem. Until then this is parked.”
caiotheodoro · GitHub · Aug 21, 2026 - 5
“# Daily ironclaw failure taxonomy — 2026-09-24 ## Suites analyzed - [officeqa (38 non-pass)](https://nearai.github.io/benchmarks/#/runs/ironclaw/officeqa/69f799a5-16f3-4577-b008-89c48b9deb82) — All 38 non-pass tasks are genuine model-quality failures by deepseek-v4-flash over OCR-digitized Treasury Bulletins; nothing in ironclaw fails. The dominant mode (recurring from the prior run) is numeric extraction/analysis error: the agent's read/grep/shell/python tool calls all succeed and return the c…”
pranavraja99 · GitHub · Sep 24, 2026 - 6
“## Objetivo Reducir el tiempo real de calificación, digitalización y generación de presentaciones sin sacrificar calidad, trazabilidad ni perder solicitudes en curso. ## Evidencia de producción (últimos 14 días) - Calificación completa: p50 157 s; p95 516 s. - Digitalización: p50 189 s. - Presentación reciente: 366 s. - DeepSeek V4 Flash Vision Exp en calificación visual: ~9–13 s por llamada exitosa. - Presentaciones con ese modelo: ~95 s por llamada y hasta tres llamadas por regeneración/revis…”
Andres-back · GitHub · Sep 4, 2026 - 7
“Thank you for sharing the public recipe *DeepSeek-v4-Flash-DSpark-2x-DGX-Spark*. I adapted your approach to my hardware setup, revalidated the configuration, and confirmed that the 200K-token context window is reproducible and reliable. **Validation Results** Starting from your recipe (your 14 runtime patches and vision overlay are used as-is), I ran three test suites on a freshly wiped pair of nodes (containers and config removed, weights reused): - **Needle-in-a-Haystack**: 9/9 pass (3 depths…”
tenaiaiai · GitHub · Sep 3, 2026 - 8
“（以下为deepseekv4flash给出的建议，提交者仅是在使用过程中遇到了标题问题而无法正确调用deepseek-v4-flash-vision-exp模型） 环境：Windows，使用编译好的 tokenicode.exe，配合 CC Switch 切换模型。 复现步骤： 打开 设置 → Provider → 模型映射 在映射框输入模型名 deepseek-v4-flash-vision-exp 输入后值被自动规范化成 deepseek-v4-flash，-vision-exp 后缀丢失 原因分析： 中的 normalizeProviderModelName 和 normalizeDeepSeekModelName 使用子串匹配来识别 DeepSeek 模型： 复制 if (compact.includes('deepseekv4flash')) return DEEPSEEK_V4_FLASH; if (compact.includes('deepseekv4pro')) return DEEPSEEK_V4_PRO; deepseek-v4-flash-vision-exp…”
Kyidella · GitHub · Sep 10, 2026 - 9
“DeepSeek V4.1 Flash 已同步上线 DeepSeek API，原生支持多模态，将模型名称更改为 deepseek-flash 即可调用最新的 V4.1 Flash 模型。旧版本模型 V4 Flash 与 V4 Flash Vision Exp 现已下线，出于兼容考虑，模型名 deepseek-v4-flash、deepseek-v4-flash-vision-exp 将被暂时路由到 V4.1 Flash。”
Danborad · GitHub · Sep 10, 2026 - 10
“Qwen3.8-27B is a 27B dense hybrid-attention model with a vision tower and an MTP draft head. It fits one H200 GPU at the full 262144 context, so it queues faster than any recipe that needs a whole node. | | Measured | | --- | --- | | One caller | 122.1 tok/s | | Concurrency 1024 | 3957.3 tok/s, still rising | | Context | 262144, the checkpoint maximum | | KV cache | 1,804,253 tokens, 6.88 full-length requests at once | | Hardware | 1 H200 GPU, TP1, no NCCL | Scope - New recipe at recipes/Qwen3.…”
mmshad · GitHub · Aug 26, 2026 - 11
“Rechercheergebnis, damit der Kandidat nicht verloren geht und beim Merge des Upstream-PR wieder hochkommt. **Heute kein Steckplatz.** ## Was es ist PrismML, veröffentlicht 2026-09-17, Apache 2.0, Basis Qwen3.8-27B. Ternär {−1, 0, +1} mit g128-FP16-Gruppenskalen. | | | |---|---| | Parameter | 27,36 B (24,35 B Backbone + 2,54 B Embed/Head + 0,46 B Vision-Turm) | | PTQ1_0 | 1,75 bpw → **5,95 GB** | | PQ2_0 | 2,13 bpw → 7,21 GB | | Vision-mmproj | + 0,63 GB | | Kontext | 262 K, multimodal, Denken p…”
ebongard · GitHub · Sep 23, 2026 - 12
“Found while configuring the new Groq platform row in prod (#1819 / PR #1825). ## What happens Promoting `Groq vision (ID extraction)` (`01M1R11M9ZPBMZZVMJ5XPQG5BQ`, `qwen/qwen3.8-27b`) to default and running a real extraction through `POST /admin/people/id-extraction`, prod logs: The ladder then fell through to BazaarLink, which answered correctly in **8.0s** — so this is contained, not user-visible. It is a wasted call on a rung that can never succeed as configured. ## What is NOT the cause Th…”
saranshisatgit · GitHub · Sep 5, 2026 - 13
“## Who As a PairPocket user on the Gemini free tier, I want the app to prefer `gemini-3.6-flash` and automatically use `gemini-3.5-flash-lite` while 3.6 is quota-blocked, so that onboarding and receipt scans keep working without waiting on a failed 3.6 call every time. ## Why Free-tier limits are per project and per model row (RPM, TPM, RPD). RPD resets at midnight Pacific Time. RPM/TPM use a rolling window. Hard-coding quota numbers is unreliable; AI Studio shows live limits. Trying 3.6 on eve…”
minsikpaul92 · GitHub · Jul 28, 2026 - 14
“1. Gemini tends to reject whole_pdf OCR pretty frequently, switching to "image" mode helps get more documents OCRed without gettng flagged. 2. Despite #1, gemini will often reject the OCR text when "paperless-gpt-auto" is used (for the second LLM pass) to try and classify the scan 3. The v0.27 logs show "googleai GenerateContent API returned a candidate with no content parts" for whatever the prescribed number of empty worker responses is set by env. 4. After the retries, the paperless-gpt work…”
phaset · GitHub · Jul 28, 2026
Rankings synthesized from community evidence and open benchmarks. See methodology. Not driven by vendor marketing.