オープンソース・開発ツール運営によるシード投票
Meta Llama vs DeepSeek vs Alibaba Cloud Qwen vs Mistral AI Mistral
“Which open-weight model family do you deploy?”
比較
- 候補1
Meta Llama
AIによる要約 — 出典をご確認ください公式ページ詳細を見る閉じる
Llama
- Model families
- Llama 4, Llama 3, Llama 2
- Llama 4 models
- Scout, Maverick
- Llama 4 architecture
- Mixture-of-experts, natively multimodal
- Scout parameters
- 17B active / 109B total, 16 experts
- Maverick parameters
- 17B active / 400B total, 128 experts
- Scout context window
- 10M tokens
- Scout hardware
- Single H100 GPU (Meta's claim)
- Access channels
- Download, Hugging Face, Kaggle
- License
- Llama 4 License and Llama 4 Use Policy
- Llama 3 range
- 3.1, 3.2, 3.3 text models
- Llama is Meta's open-model line; the developer site now leads with the Muse model family and Meta Model API.
- Meta's $0.19/Mtok (3:1 blended) figure for Llama 4 Maverick is a cost estimate assuming distributed inference, not a…
- Llama 4 Scout is described by Meta as fitting a single H100 GPU; no other hardware requirements are published on the…
出典: Llama（公式） · Llama Docs: Get started with Llama (dev.meta.ai)（公式） · Llama 4 model page (dev.meta.ai)（公式） · Meta Model API home (dev.meta.ai)（公式） · The Llama 4 herd blog post (ai.meta.com)（公式）
2026-10-30に確認
- 候補2
DeepSeek
AIによる要約 — 出典をご確認ください公式ページ詳細を見る閉じる
DeepSeek
- Consumer access
- Web chat, iOS and Android app, Harness desktop
- Consumer price
- Free (official app page)
- deepseek-flash input (miss)
- CNY 1 off-peak / 2 peak per 1M tokens
- deepseek-flash output
- CNY 4 off-peak / 8 peak per 1M tokens
- deepseek-v4-pro input (miss)
- CNY 4.5 off-peak / 9 peak per 1M tokens
- deepseek-v4-pro output
- CNY 13.5 off-peak / 27 peak per 1M tokens
- Context length
- 1M tokens (both models)
- Max output
- 384K tokens (both models)
- Search and Fetch APIs
- CNY 5 per 1,000 successful requests each
- Image understanding
- deepseek-flash only; V4-Pro not supported
- Official pricing lists API prices in CNY per 1M tokens; off-peak is half the peak price…
- Search and Fetch calls made through DeepSeek Harness, Claude Code or Codex are unbilled until 00:00 Beijing time on 1…
- deepseek-v4-flash is a legacy name that now routes to the deepseek-flash model at Flash prices.
出典: DeepSeek（公式） · DeepSeek homepage (English)（公式） · DeepSeek Download page (English)（公式） · DeepSeek API Docs - Models & Pricing (zh-cn)（公式）
2026-10-30に確認
- 候補3
Alibaba Cloud Qwen
AIによる要約 — 出典をご確認ください公式ページ詳細を見る閉じる
Qwen
- Free access
- Free to use (qwen.ai homepage)
- Platforms
- Web, iOS, Android, macOS, Windows
- Main features
- Thinking, search, Deep Research, Web Dev, Artifacts
- Image and video
- Image Edit, Image and Video Generation
- API format
- OpenAI API compatible
- API price: Qwen3.8-Max
- $2 input / $6 output per 1M tokens (Singapore)
- API price: Qwen3.7-Plus
- $0.4 input / $1.6 output per 1M tokens (up to 256K)
- API free quota
- 1M tokens per model, 90 days (Singapore only)
- The qwen.ai homepage describes Qwen as free to use and lists the Qwen API platform and apps for web, phone and desktop.
- Alibaba Cloud Model Studio bills API calls pay-as-you-go by default; the listed prices are standard prices and vary by…
- Free API quota applies only in Singapore, per the official pricing page.
出典: Qwen（公式） · Qwen home (official)（公式） · Alibaba Cloud Model Studio model inference pricing（公式）
2026-10-30に確認
- 候補4
Mistral AI Mistral
AIによる要約 — 出典をご確認ください公式ページ詳細を見る閉じる
Mistral
- Free plan
- Yes, limited access to Vibe; includes Mistral Studio
- Pro plan
- $14.99/mo excluding taxes
- Student plan
- $5.99/mo for verified students, max 12 months
- Team plan
- $24.99/user/mo excluding taxes
- Enterprise plan
- Custom pricing, contact sales
- Platforms
- Web, iOS, Android, CLI, IDE (VS Code, JetBrains)
- API pricing example
- Mistral Large $0.5 in / $1.5 out per M tokens
- Self-hosting
- Yes; open-weight models, Apache 2.0 for research/individual use
- The pricing page covers the Vibe chat and coding plans plus API pricing; USD is the default view with a EUR option.
- Batch processing cuts API prices by 50%; cached input tokens can cost up to 90% less for repeated prompts.
- Commercial self-hosted deployments require a Mistral license with separate terms, per the official FAQ.
出典: Pricing | Mistral (official pricing page)（公式）
2026-10-30に確認
投票するとすぐに結果が表示されます。他の人には投票終了後に表示されます。