Kimi K2.5 Review: Moonshot AI's Open Model with Agent Swarms
Kimi K2.5 deploys up to 100 parallel sub-agents, scores 50.2% on HLE with Tools, and costs $0.60/$3.00 per million tokens. Full review with benchmarks.
Moonshot AI released Kimi K2.5 on January 27, 2026, and it immediately set a record no other model has matched: 50.2% on Humanity's Last Exam with Tools enabled: the best score posted by any model on one of AI's hardest benchmarks (Kimi K2.5 Tech Blog). It also scored 78.4% on BrowseComp Agent Swarm, beating GPT-5.2 Pro.
The model's defining feature is its Agent Swarm mode: up to 100 parallel sub-agents running simultaneously, handling up to 1,500 tool calls at once, with task execution 3–4.5x faster than single-agent baselines. As of February 2026, Kimi K2.5 is the most capable open-weights model for autonomous multi-step agentic work.
This review covers the architecture, Agent Swarm in depth, benchmarks, writing quality, multimodal capabilities, pricing, and practical use cases.
What is Kimi K2.5?
Kimi K2.5 is Moonshot AI's second-generation open-weights multimodal language model, released under a modified MIT license that allows commercial use with attribution. It processes text, images, and video natively, not through adapter layers or separate models.
Architecture overview:
- Parameters: 1 trillion total, ~32B active (MoE)
- Context window: 256,000–262,000 tokens
- Modalities: Text, image, video (native, not adapter-based)
- License: Modified MIT (commercial use with attribution)
- Weights: Available on Hugging Face (moonshotai/Kimi-K2.5)
The 1T total parameter count with only 32B active is an unusually sparse MoE design. The sparsity enables very low inference cost per token while the large expert pool provides broad knowledge coverage. This is the architecture decision that allows Kimi K2.5 to run at $0.60/M input tokens while competing with models that cost 5–10x more.
The 256K context window is one of the largest among open-weights models, enabling genuine long-document workflows, full codebases, legal discovery sets, book-length research documents, in a single request.
Agent Swarm capability
Agent Swarm is the feature that most distinguishes Kimi K2.5 from every other open-weights model. It is currently in research preview and available through the Kimi API.
How it works
When Agent Swarm mode is activated, Kimi K2.5 decomposes a complex task into subtasks and dispatches each to a dedicated sub-agent. Sub-agents operate in parallel, each capable of using tools independently. The orchestrating model then aggregates results and synthesizes a final output.
Capacity figures from the official announcement:
- Up to 100 parallel sub-agents
- Up to 1,500 simultaneous tool calls across the swarm
- 3–4.5x faster task execution compared to single-agent mode
- 76% lower cost than Claude Opus 4.5 for equivalent swarm tasks
The practical effect on a benchmark like HLE (Humanity's Last Exam) is dramatic. Single-agent HLE performance for frontier models typically sits in the 20–35% range. With Agent Swarm enabled, allowing the model to research, verify, and cross-check answers in parallel, Kimi K2.5 reaches 50.2%, higher than any other model tested (InfoQ).
BrowseComp Agent Swarm
BrowseComp measures a model's ability to answer hard, multi-hop questions by searching the web. With Agent Swarm enabled, Kimi K2.5 scores 78.4%, above GPT-5.2 Pro on the same task. This is a concrete demonstration that parallel search agents can outperform a single more-capable model on research tasks.
Wall-clock time reduction
The 3–4.5x speed improvement reflects real-world workflow acceleration, not just a benchmark number. A research task that takes a single agent 10 minutes can complete in 2–3 minutes with swarm parallelism. For development pipelines and automated research workflows where latency compounds, this is meaningful.
Limitations of Agent Swarm
Agent Swarm is still research preview. It adds orchestration complexity compared to single-model API calls and requires understanding the Kimi API's swarm-specific parameters. Teams building on it should budget for prompt engineering time to structure tasks in ways that decompose cleanly into parallel sub-tasks.
Benchmark performance
Coding
HumanEval: 99.0%: effectively perfect on this coding benchmark. This does not mean the model writes perfect production code, but it can solve standard algorithm problems and implement well-defined functions reliably.
SWE-Bench Verified: 76.8%: strong performance on real GitHub issue resolution, behind MiniMax M2.5 (80.2%) but ahead of most other models in its class.
Science and reasoning
GPQA Diamond: 87.6%: the highest GPQA score among the major Chinese open-weights models reviewed here, and competitive with the best proprietary frontier models.
MMLU: 92.0%: strong on the standard general knowledge benchmark.
Research and agentic tasks
HLE with Tools (full swarm): 50.2%: the benchmark-defining number. Humanity's Last Exam is a set of extremely hard questions across domains assembled to resist AI solving. A 50% score with tools is the frontier of what any model can currently do.
BrowseComp Agent Swarm: 78.4%: above GPT-5.2 Pro.
Chatbot Arena
Chatbot Arena rating: 1447. This is a human preference score from LMSYS's live arena where models are compared in blind A/B tests. A rating of 1447 places Kimi K2.5 among the top-tier conversational models, which matters for consumer-facing applications where raw benchmark scores do not tell the whole story.
Artificial Analysis Intelligence Index
Score of 47, ranked #2 among open-weights models behind GLM-5 on the AA index at time of publication.
Writing quality
Kimi K2.5 is notable among Chinese open-weights models for its writing quality, something that has historically been a gap between Chinese-developed models and their Western counterparts.
The lechmazur creative writing benchmark ranks Kimi K2.5 at position #16 globally, with a score of 8.068. Its predecessor, K2-0905, holds position #7 with a score of 8.331, the best creative writing score ever recorded for a Chinese-developed model.
For context: most Chinese open models rank in the 20–50 range on lechmazur. Kimi at #16 is genuinely competitive with strong Western models and significantly ahead of MiniMax M2.5 (which ranked 19th for the M2.1 predecessor) and GLM-5 (which has not posted strong writing rankings).
This matters for:
- Creative writing assistants where output quality is directly evaluated by users
- Marketing and content generation where prose style affects engagement
- Conversational AI products where naturalness of dialogue affects retention
- Educational tutoring where explanations benefit from clear, readable writing
If you are building any product where the quality of generated prose is evaluated by end users, Kimi K2.5 is the correct Chinese open-weights model to start with.
Vision and multimodal
Kimi K2.5 processes images and video natively, not via a bolted-on vision adapter. This is architecturally important, native multimodal models generally show better cross-modal reasoning than models where visual and language components are trained separately and fused at inference time.
Practical multimodal capabilities:
- Image understanding: Read charts, diagrams, screenshots, documents, and photographs
- Video analysis: Process video frames to understand sequences of events
- Document processing: Handle PDFs with embedded images or mixed-format documents
- Visual code analysis: Analyze UI screenshots and generate corresponding code
The native multimodality combined with the 256K context window and Agent Swarm makes Kimi K2.5 genuinely capable for complex agentic workflows that involve visual inputs, for example, a research agent that browses websites (seeing screenshots), reads documents (PDFs with images), and synthesizes findings in parallel.
Among open-weights models in the frontier tier, native multimodality at this performance level is still rare. GLM-5 and MiniMax M2.5 are primarily text models. Kimi K2.5's vision capability is a real differentiator.
Pricing
| Tier | Price per 1M tokens |
|---|---|
| Input (cache miss) | $0.60 |
| Input (cache hit) | $0.10 |
| Output | $3.00 |
Comparison to proprietary alternatives:
| Model | Input | Output |
|---|---|---|
| Kimi K2.5 | $0.60 | $3.00 |
| Claude Opus 4.5 | $5.00 | $25.00 |
| GPT-5.2 Pro | ~$5.00+ | ~$20.00+ |
| GLM-5 | $1.00 | $3.20 |
| MiniMax M2.5 | $0.30 | $1.20 |
Against Claude Opus 4.5, Kimi K2.5 is approximately 9x cheaper on input and output (eesel.ai). The Agent Swarm capability specifically runs at 76% lower cost than Claude Opus 4.5 for equivalent tasks, a figure that matters because the benchmark comparison is apples-to-apples (both in agentic mode).
The cache hit price of $0.10/M makes repeated large-context requests very cheap. For applications that repeatedly reference a large document or codebase, code review agents, document analysis pipelines, the effective cost per query drops substantially once the context is cached.
Use cases
Autonomous research agents
This is Kimi K2.5's clearest strength. The combination of Agent Swarm (100 parallel sub-agents), 78.4% BrowseComp performance, and 256K context creates a model that can autonomously research complex questions, synthesize findings from multiple sources, and deliver results faster than any other open-weights option. For teams building research automation tools, literature review systems, or competitive intelligence pipelines, K2.5 sets the bar.
Complex agentic pipelines
The 1,500 simultaneous tool call capacity means Kimi K2.5 can orchestrate deeply parallel workflows, multiple API calls, database queries, and computation steps happening simultaneously. This is the right model for developers building systems where a single user request triggers many downstream automated actions.
Multimodal document processing
Native video and image understanding combined with a 256K context window enables workflows that mix text, screenshots, PDFs with embedded diagrams, and video transcripts in a single request. Customer service systems, compliance review tools, and educational platforms that handle mixed media benefit from this.
Creative and long-form writing
At #16 on lechmazur, Kimi K2.5 is the highest-quality Chinese open-weights model for prose generation. For content platforms, marketing automation, or creative writing tools where output quality is evaluated by end users, K2.5 is the correct choice among open models in this price range.
High-difficulty reasoning tasks
The 87.6% GPQA Diamond and 99.0% HumanEval scores make Kimi K2.5 the strongest open-weights model for hard science and engineering problems. Graduate-level tutoring, research assistant tools, and scientific computing support are natural applications.
Cost-sensitive production deployments
At $0.60/$3.00 with strong general performance, Kimi K2.5 is priced for production scale. The $0.10/M cache hit price makes it especially attractive for applications that reuse large contexts, system prompts, retrieved documents, codebase context, across many queries.
Limitations
Agent Swarm is research preview
The 100-agent parallel execution mode is not yet stable for production use. Teams building on it should expect API changes and should architect their pipelines to degrade gracefully to single-agent mode.
Modified MIT license requires attribution
Unlike GLM-5 (pure MIT), Kimi K2.5's modified MIT license requires attribution when the weights are incorporated into products. A minor constraint for most commercial uses but worth reviewing for specific deployment contexts.
SWE-Bench is not the leader
Despite the impressive headline numbers, Kimi K2.5 scores 76.8% on SWE-Bench Verified, below MiniMax M2.5 (80.2%) for autonomous code repair tasks. For coding-first applications where SWE-Bench performance is the primary success metric, M2.5 has an edge.
Active parameter count is very sparse
32B active out of 1T total is an extremely sparse ratio. This works well on benchmarks, but some developers report uneven quality on tasks that fall outside the model's strongest domains. Running your own evals on your specific task distribution before committing to production is good practice.
Availability during peak demand
Moonshot AI became China's fastest-growing decacorn in early 2026, with 20-day post-launch revenue exceeding all of 2025 (The China Academy). High demand has caused intermittent capacity constraints on the API. Self-hosting on your own infrastructure using the open weights is the hedge against this.
Verdict
Kimi K2.5 is the most capable open-weights model for agentic applications as of February 2026. The Agent Swarm system, 100 parallel sub-agents, 1,500 simultaneous tool calls, 3–4.5x execution speed, is a genuine architectural capability that no other open model matches. The 50.2% HLE with Tools score and 78.4% BrowseComp swarm result back that up with hard numbers.
Beyond the agent story, K2.5 is also the best open-weights Chinese model for writing quality (#16 globally on lechmazur), offers native multimodality, and comes with a 256K context window, all at $0.60/$3.00 per million tokens, roughly 9x cheaper than Claude Opus 4.5.
The tradeoffs are real: MiniMax M2.5 is better for pure coding tasks, GLM-5 is better calibrated on factual accuracy, and Agent Swarm is still in preview. But if you are building autonomous research tools, complex multi-step pipelines, or multimodal applications and want open weights, Kimi K2.5 is the model to evaluate first.
Available on AfricanAI: run Kimi K2.5 directly in your browser alongside other frontier models.
Sources: