Model Drop

Kimi K2.5

January 27, 2026 2 min read

Kimi K2.5, which Moonshot AI released on January 27, 2026, is an open-weights model that adds native image and video input to the Kimi K2 line, along with a self-directed agent swarm mode.

Model: Kimi K2.5 (moonshotai/Kimi-K2.5)
Maker: Moonshot AI
Ship date: January 27, 2026
Pricing: $0.60 / $3 per M tokens (Moonshot API); free weights
Context: 256K tokens
Weights: Open (Modified MIT)
Where you can use it:
- Kimi web and app, with Agent Swarm in beta
- Moonshot API, Kimi Code, Hugging Face
When to reach for it:
- Agentic search and long tool-calling runs
- Front-end code from images or video
- Skip it for the hardest coding work
How it compares:
- Leads open peers DeepSeek V3.2 and GLM-4.7
- Near Gemini 3 Pro, short of Claude Opus 4.5
What makes it different:
- Native vision, trained on ~15T mixed tokens
- Agent swarm of up to 100 sub-agents
- One model with instant and thinking modes
Early reactions: Artificial Analysis called it the new leading open weights model, and early users quoted by Zvi Mowshowitz put it near Gemini 3 level but a bit short of Claude Opus 4.5 on polish.
Read more: Kimi K2.5 Tech Blog: Visual Agentic Intelligence and the model card on Hugging Face

The benchmarks

Moonshot ran K2.5 in thinking mode at temperature 1.0, top-p 0.95 and 256K context against GPT-5.2 at xhigh effort, Claude Opus 4.5 with extended thinking, Gemini 3 Pro at a high thinking level and DeepSeek V3.2 in thinking mode, plus Qwen3-VL-235B-A22B-Thinking on vision; rival scores marked as reruns in Moonshot's table are its own, GPT-5.2 failed about 10% of vision items (counted wrong), SWE-bench and Terminal-Bench 2.0 scores come from non-thinking mode, and HLE is the full text-and-image set. With context management, Moonshot's BrowseComp score rises to 74.9%.

Moonshot AI's numbers

Humanity's Last Exam (with tools) (full set, expert questions): 50.2%. +4.7 over GPT-5.2 (45.5%), +7.0 over Claude Opus 4.5 (43.2%), +4.4 over Gemini 3 Pro (45.8%).
Humanity's Last Exam (no tools) (full set, expert questions): 30.1%. −4.4 under GPT-5.2 (34.5%), −0.7 under Claude Opus 4.5 (30.8%), −7.4 under Gemini 3 Pro (37.5%).
BrowseComp (agentic web research): 60.6%. −5.2 under GPT-5.2 (65.8%), +23.6 over Claude Opus 4.5 (37.0%), +22.8 over Gemini 3 Pro (37.8%), +9.2 over DeepSeek V3.2 (51.4%).
SWE-bench Verified (real GitHub fixes): 76.8%. −3.2 under GPT-5.2 (80.0%), −4.1 under Claude Opus 4.5 (80.9%), +0.6 over Gemini 3 Pro (76.2%), +3.7 over DeepSeek V3.2 (73.1%).
Terminal-Bench 2.0 (command-line agent tasks): 50.8%. −3.2 under GPT-5.2 (54.0%), −8.5 under Claude Opus 4.5 (59.3%), −3.4 under Gemini 3 Pro (54.2%), +4.4 over DeepSeek V3.2 (46.4%).
MMMU-Pro (multimodal reasoning): 78.5%. −1.0 under GPT-5.2 (79.5%), +4.5 over Claude Opus 4.5 (74.0%), −2.5 under Gemini 3 Pro (81.0%), +9.2 over Qwen3-VL-235B-A22B-Thinking (69.3%).
Video-MMMU (learning from video): 86.6%. +0.7 over GPT-5.2 (85.9%), +2.2 over Claude Opus 4.5 (84.4%), −1.0 under Gemini 3 Pro (87.6%), +6.6 over Qwen3-VL-235B-A22B-Thinking (80.0%).
GPQA Diamond (grad-level science): 87.6%. −4.8 under GPT-5.2 (92.4%), +0.6 over Claude Opus 4.5 (87.0%), −4.3 under Gemini 3 Pro (91.9%), +5.2 over DeepSeek V3.2 (82.4%).

Independent numbers

Artificial Analysis Intelligence Index (current v4.3.2, AA estimate): 23.5. +2.0 over DeepSeek V3.2 (21.5), +1.3 over GLM-4.7 (22.2), +2.6 over MiniMax-M2.1 (20.9), −5.6 under Claude Opus 4.5 (29.1).
GDPval-AA (agentic knowledge work Elo): 1309.
AA-Omniscience hallucination rate (lower is better): 64%. −10 under Kimi K2 Thinking (74%).