Table of Contents
April 2026 LLM leaderboard snapshot — what to pick by task. Frontier APIs (Claude 3.7 Opus, GPT-5) lead the hardest categories; open-weight models (Llama 3.3 70B, DeepSeek V3, Qwen 2.5 72B) close most gaps.
Frontier API: Claude 3.7 Opus, GPT-5, Gemini 2.0 Pro. Open-weight 70B+: Llama 3.3 70B, Qwen 2.5 72B, DeepSeek V3. Open-weight mid: Mistral Small 3 24B, Qwen 2.5 32B. Open-weight small: Mistral 7B v0.3, Llama 3.1 8B, Qwen 2.5 7B. Reasoning: o3, R1-Distill family. Code: Qwen 2.5 Coder 32B, Codestral, GPT-5.
English
- Frontier: Claude 3.7 Opus, GPT-5, Gemini 2.0 Pro
- Open-weight 70B: Llama 3.3 70B (most-deployed), Qwen 2.5 72B
- Open-weight 24-32B: Mistral Small 3, Qwen 2.5 32B
- Open-weight 7-8B: Mistral 7B v0.3, Llama 3.1 8B
Multilingual
- Chinese / Japanese / Korean: Qwen 2.5 family
- European languages: Mistral Large 2, Llama 3.3 70B
- Low-resource: Aya 32B (Cohere) or hosted Google Translate
Specialised
- Reasoning: OpenAI o3, R1-Distill family
- Code: Qwen 2.5 Coder 32B, Codestral 22B, GPT-5
- Vision: Qwen2-VL 72B, GPT-5, Gemini 2.0 Pro
- Long-context: Llama 3.1 128K, Gemini 2.0 1M
Verdict
Pick by task + cost / residency profile. Frontier APIs lead the hardest tasks; open-weight covers ~95% of production well. The hybrid pattern (open-weight bulk + frontier API for hardest 5-10%) is the dominant production architecture.
Bottom line
Hybrid pattern: open-weight + frontier fallback. See benchmark rankings.
