Strong chart understanding. +10 pts over Opus 4.6 (78.9). 3x higher vision resolution helps document/chart tasks.

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

Tool-integrated chart reasoning: +2.04% over Qwen2.5-VL-7B on CharXiv Avg via image cropping and code-based computation with agentic RL.

Chart reasoning SOTA. Beats GPT-5.4 (82.8), Gemini 3.1 Pro (80.2), Claude Opus 4.6 (65.3) per Meta announcement.

CharXiv Reasoning Quality.

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

Highest among compared models. Up from Gemini 3 Flash (80.3).

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

Below Gemini 3.5 Flash (84.2). Above most peers.

Below Gemini 3.5 Flash (84.2). Above Sonnet 4.6 (72.4).

3B tool-integrated chart model: +5.04% over Qwen2.5-VL-3B on CharXiv Avg, using DuoChart data and agentic RL with cropping + code tools.

Chart understanding and reasoning. Between Claude Opus 4.6 (78.9) and Muse Spark (86.4).

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

Without python tool. With python: 86.7. Native multimodal chart understanding.

Below Gemini 3.5 Flash (84.2).

With tools. Chart reasoning on arxiv figures.

Near Qwen 3.5 397B MoE on document understanding. Beats Claude Opus 4.5.

Chart reasoning; below Fable 5 (86.5) but above GLM-5.2. Encoder-free vision architecture.

MoE 3B active/35B total. Matches Sonnet 4.5 on document chart understanding.

CharXiv RQ.

Below all frontier models. Gemini 3.5 Flash leads (84.2).

Strong document understanding. Beats GPT-5 mini (68.6) but trails Qwen3.5 27B (79.5).

No tools. Chart reasoning on arxiv figures.