Highest among compared models. Up from Gemini 3 Flash (81.2).

With tools. 1pt over GPT-5.4 (82.1).

Multimodal reasoning no-tools; edges GPT-5.5 (81.2%); 84.6% with tools.

Visual understanding and reasoning. Without tool use, reasoning effort xhigh.

No tools. Matches GPT-5.4, above Gemini 3.1 Pro (80.5).

Score as cited in Kimi K3 blog (source: Anthropic official). Fable 5 hit fallbacks on 35% of tasks.

Multimodal reasoning at Flash-tier cost. Competitive with Gemini 3 Pro.

Terra nearly matches GPT-5.5 (81.2%) on MMMU Pro.

Multimodal understanding without tools. Strong visual reasoning.

Second-best MMMU Pro score, behind Gemini 3.1 Pro Preview (82.4%). Strong multimodal showing for Meta Superintelligence Labs debut model.

Native multimodal model matching Gemini 3 Flash (81.2) on MMMU Pro.

Native vision-language model. Strong multimodal reasoning.

Score as cited in Kimi K3 blog (source: Anthropic official).

Strong multimodal reasoning for an open-weight model.

Luna trails GPT-5.5 (81.2%) on MMMU Pro.

Multimodal vision benchmark. Up from 49.7% on Gemma 3 27B.

Budget-tier ($0.25/1M input) yet competitive on multimodal benchmarks.

Strong multimodal reasoning for 27B.

Matches Gemma 4 31B (76.9%). Strong vision-language reasoning for a MoE model.

Below Gemini 3.5 Flash (83.6). Above Sonnet 4.6 (74.5).

Multimodal understanding. Strong for a small model — competitive with some frontier scores from late 2025.

Below all frontier models. Gemini 3.5 Flash leads (83.6).

MoE multimodal. 31B dense reaches 76.9%.

Vision reasoning; below Gemini 3.5 Flash (83.6) but strong for open-weights multimodal.

Matches Inkling on MMMU Pro.

9B params. Outperforms Gemini 2.5 Flash-Lite (59.7) on visual reasoning. Strong for its size class.

Vision benchmark. Strong multimodal reasoning.

Ternary variant. Vision tower ships in 4-bit; multimodal agentic capability retained on-device.

Beats GPT-5 mini (67.3). Strong for an open-weight 33B VLM.

1-bit variant. Largest category drop (vision 72.6→59.6) — the 1-bit footprint trade-off lands heaviest on vision tasks.

+3.9 over Qwen3-VL-8B-Instruct base. Open RL recipe with 600K samples across 59 datasets.

Official model card score. 17B active params, 128 experts.

Multimodal understanding from a 4B model. Strong for edge deployment.

Official model card score. 17B active params, 16 experts, 10M context.

Multimodal reasoning at 2.3B active params. Natively multimodal from pretraining.