HELM
Indexed results from HELM. Each ranking retains its own metric and evaluation conditions.
HELM GPQA
helm · HELM Capabilities · chain_of_thought_correctness · percent
- gemini-3-pro-previewgoogle80.269%
- Qwen3-235B-A22B-Instruct-2507-FP8qwen72.646%
- OpenAI: gpt-oss-120bopenai68.386%
- Claude Opus 4anthropic66.592%
- DeepSeek: R1 0528deepseek66.592%
- Writer: Palmyra X5writer66.143%
- MoonshotAI: Kimi K2 0711moonshotai65.247%
- Llama-4-Maverick-17B-128E-Instruct-FP8meta-llama65.022%
- Claude Sonnet 4anthropic64.35%
- Qwen: Qwen3 Next 80B A3B Thinkingqwen63.004%
- OpenAI: gpt-oss-20bopenai59.417%
- GLM-4.5-Air-FP8z-ai59.417%
HELM IFEval
helm · HELM Capabilities · ifeval_strict_accuracy · percent
- Claude Opus 4anthropic91.774%
- Llama-4-Maverick-17B-128E-Instruct-FP8meta-llama90.758%
- gemini-3-pro-previewgoogle87.646%
- MoonshotAI: Kimi K2 0711moonshotai85.028%
- Claude Sonnet 4anthropic83.888%
- OpenAI: gpt-oss-120bopenai83.611%
- Qwen3-235B-A22B-Instruct-2507-FP8qwen83.518%
- DeepSeek: DeepSeek V3deepseek83.241%
- Writer: Palmyra X5writer82.286%
- OpenAI: GPT-4o (2024-11-20)openai81.731%
- Amazon: Nova Pro 1.0amazon81.547%
- GLM-4.5-Air-FP8z-ai81.177%
HELM MMLU-Pro
helm · HELM Capabilities · chain_of_thought_correctness · percent
- gemini-3-pro-previewgoogle90.3%
- Claude Opus 4anthropic85.9%
- Qwen3-235B-A22B-Instruct-2507-FP8qwen84.4%
- Claude Sonnet 4anthropic84.3%
- MoonshotAI: Kimi K2 0711moonshotai81.9%
- Llama-4-Maverick-17B-128E-Instruct-FP8meta-llama81%
- Writer: Palmyra X5writer80.4%
- OpenAI: gpt-oss-120bopenai79.5%
- DeepSeek: R1 0528deepseek79.3%
- Qwen: Qwen3 Next 80B A3B Thinkingqwen78.6%
- GLM-4.5-Air-FP8z-ai76.2%
- OpenAI: gpt-oss-20bopenai74%
HELM Omni-MATH
helm · HELM Capabilities · omni_math_accuracy · percent
- Qwen3-235B-A22B-Instruct-2507-FP8qwen71.833%
- OpenAI: gpt-oss-120bopenai68.833%
- MoonshotAI: Kimi K2 0711moonshotai65.4%
- OpenAI: gpt-oss-20bopenai56.467%
- gemini-3-pro-previewgoogle55.55%
- Claude Sonnet 4anthropic51.25%
- Claude Opus 4anthropic51.117%
- Google: Gemini 2.5 Flash Litegoogle47.967%
- Qwen: Qwen3 Next 80B A3B Thinkingqwen46.683%
- DeepSeek: R1 0528deepseek42.4%
- Llama-4-Maverick-17B-128E-Instruct-FP8meta-llama42.233%
- Writer: Palmyra X5writer41.45%
HELM WildBench
helm · HELM Capabilities · wildbench_score_rescaled · percent
- Qwen3-235B-A22B-Instruct-2507-FP8qwen86.626%
- MoonshotAI: Kimi K2 0711moonshotai86.187%
- gemini-3-pro-previewgoogle85.872%
- OpenAI: gpt-oss-120bopenai84.494%
- Claude Opus 4anthropic83.3%
- DeepSeek: DeepSeek V3deepseek83.052%
- OpenAI: GPT-4o (2024-11-20)openai82.835%
- DeepSeek: R1 0528deepseek82.809%
- Claude Sonnet 4anthropic82.47%
- Google: Gemini 2.5 Flash Litegoogle81.769%
- Qwen: Qwen3 Next 80B A3B Thinkingqwen80.737%
- Llama-4-Maverick-17B-128E-Instruct-FP8meta-llama79.969%
Models
LoadingBenchmarks
Explore every benchmark
LoadingThese charts contain model-linked results. The charts above also include unlinked source profiles. Compare scores within each evaluation. Model quality, usage, transparency and system performance measure different things.