Kimi K3 Review: Moonshot's 2.8T Open Frontier Model

Moonshot AI dropped Kimi K3 on July 16, 2026 — a 2.8-trillion-parameter open-weight model that immediately became the largest openly available language model in existence. It tops LMArena’s Frontend Code Arena, scores 93.5% on GPQA Diamond, and ships with a 1-million-token context window. Weights drop on July 27.

This is the successor to the Kimi K2 family (K2, K2.5, K2.6, K2.7 Code), and it’s a meaningful jump.

Ease of Use: 8/10

Accessing K3 is straightforward but fragmented. You have four options:

Kimi Code CLI (easiest) — pip install kimi-code && kimicode gives you a terminal coding agent with background tasks, plan mode, skills, and nested agent orchestration. Version 0.26.0 (shipped with K3) adds a Goal Queue, Web Mode, and context compaction. It works like Claude Code but with K3 under the hood.

Direct API — Standard OpenAI-compatible endpoint at api.moonshot.ai/v1/chat/completions. Pricing is $3/1M input tokens and $15/1M output tokens. For coding workloads with 90% cache hit rate, effective cost drops to roughly $1.77 per 1M input + 100K output — about 60% cheaper than standard pricing.

Kimi App — Web and mobile chat interface with tiered plans from free (limited) to $199/month (Allegretto plan, 1M context, priority access).

API via Partners — Available through Together AI, Fireworks, and other inference providers at launch.

The fragmentation isn’t ideal — you choose your access path before you know which features you need. But each path is well-documented, and the OpenAI-compatible API means existing tooling works with minimal changes.

The one real friction point: the CLI requires Python 3.10+ and can be picky about dependency versions. pip install in a fresh venv is the safest route.

Features: 9/10

K3 packs an unusual combination of capabilities for an open-weight model:

The KDA attention mechanism is the architectural highlight. Moonshot claims it’s up to 6.3x faster in million-token contexts compared to standard attention. In practice, long-document processing (100K+ tokens) feels snappy — no multi-minute wait for the first token.

Training efficiency improvements (~25% gain) mean Moonshot got more out of their compute budget. The model uses Muon optimizer for large matrices and Adam elsewhere — the same hybrid approach Thinking Machines used with Inkling.

The Kimi Code CLI v0.26.0 deserves special mention. It’s a full agent harness with:

• Background tasks — kick off long-running operations and check results later
• Plan mode — the agent proposes a multi-step plan before executing
• Skills — reusable, composable workflows
• Nested agents — subagents for research, coding, and planning
• Web mode — browse and fetch context from URLs
• Context compaction — automatically summarizes old context to stay within limits

This is more polished than most open-weight agent frameworks shipping today.

Performance: 9/10

K3 benchmarks are genuinely impressive for an open model:

• GPQA Diamond: 93.5% — near Claude Fable 5 territory
• LMArena Frontend Code Arena: #1 at launch
• SWE-bench Verified: Competitive with top closed models
• 1M context retrieval: Strong needle-in-haystack at full context length

On Artificial Analysis, K3 ranks fourth overall among all models, behind only GPT-5.6 Sol, Claude Fable 5, and Gemini 3.5 Pro — all closed, proprietary systems. For an open-weight model available under Apache 2.0, this is unprecedented.

Speed is reasonable for a 2.8T MoE. The active parameters per token are much smaller (typical MoE sparsity), though Moonshot hasn’t published the exact active count. Through the API, first-token latency at short contexts is ~400-800ms. At 500K+ contexts, KDA attention keeps it under 3 seconds — notably faster than standard attention models at that length.

The one wrinkle: output quality depends on the Kimi Code CLI version. The API gives consistent results; the app’s free tier may use a quantized or distilled variant with lower quality.

Documentation: 8/10

Moonshot publishes solid documentation across three channels:

Official API docs at platform.moonshot.ai cover authentication, endpoints, streaming, tool use, and function calling. The OpenAI-compatible format means less learning curve.

Kimi Code docs are more detailed than expected — setup guides, agent configuration, skill authoring, and a cookbook with example workflows. The section on context compaction is particularly well-written.

Technical blog post at kimi.com/blog/kimi-k3 covers the architecture at a research-paper level: KDA attention, training infrastructure (NVIDIA GB300 NVL72 systems), and benchmark methodology. A full technical report is promised alongside the weight release on July 27.

The gaps: no interactive playground for prompt experimentation (unlike Thinking Machines’ Inkling Playground), and the pricing documentation buries the cache-hit discount in a footnote. New users could easily miss that their effective cost is 60% lower with caching.

Support: 7/10

Support is the weakest area. For a model targeting production use, the support channels are limited:

• GitHub Issues for Kimi Code — responsive but community-driven
• Discord community — active, mostly English-language
• Email support for paid API plans — response times vary (2-24 hours based on reports)
• No phone support and no dedicated Slack/Teams channel for enterprise customers

Enterprise agreements exist (contact sales) but there’s no published SLA below the enterprise tier. For production deployments at scale, you’ll want to negotiate a contract rather than relying on the standard API.

The community on Discord is helpful, and Moonshot engineers occasionally answer technical questions directly — but “occasionally” isn’t a support plan.

Verdict

Kimi K3 is the most capable open-weight model available at launch — beating every other open model on reasoning and coding benchmarks while matching some closed flagships. The 1M context window with KDA attention makes it the best open choice for long-context workloads. The Kimi Code CLI is a polished agent harness that rivals paid alternatives.

Use it if: you need frontier-level intelligence in an open package, want 1M+ token context processing, or want a capable coding agent without monthly subscriptions.

Skip it if: you need enterprise SLAs, or you’re on a tight budget — the $15/M output tokens is steep, and DeepSeek V4 offers ~80% of the quality at ~10% of the price.

The open-weight release on July 27 will be the real test. If the weights are genuinely Apache 2.0 and runnable on consumer hardware (quantized), K3 becomes the default open model for serious work. If the hardware requirements are prohibitive, it’ll remain an API-first product despite the open license.

Update (Jul 27): We’ll update this review when the weights land with self-hosting benchmarks and quantized performance data.

Pros

• Top-tier benchmarks for an open model, competitive with closed flagships
• 1M token context with fast KDA attention
• Polished CLI agent harness with background tasks, skills, and nested agents
• OpenAI-compatible API reduces integration friction
• Open weights (Apache 2.0) due July 27

Cons

• API pricing ($15/M output) is expensive for heavy use
• No published SLA below enterprise tier
• Customer support is community-driven, not enterprise-grade
• Access is fragmented across CLI, API, app, and partners
• Full weights not yet available (requires waiting until July 27)

Pricing

• API: $3/1M input tokens, $15/1M output tokens
• Cache-hit input: $0.30/1M tokens (typically 90% cache rate for coding workloads)
• Kimi App: Free (limited) → $19/month (Moderato, 256K context) → $199/month (Allegretto, 1M context, priority access)
• Pricing not publicly listed for enterprise contracts — contact Moonshot sales

📊 See how Kimi K3 compares → /comparisons/

Sources

• Moonshot Kimi K3 Technical Blog
• Artificial Analysis: Kimi K3 Intelligence & Pricing
• Kimi K3 Review: Benchmarks and K2 Comparison
• Kimi K3 API Pricing Explained
• Kimi K3: Specs, Pricing & Release
• Kimi Code K3: Setup, Plans and Cache Discipline

📖 Related Reads

• ToolBrain — tool reviews, LLM comparisons, and AI workflow guides
• CodeIntel Log — code quality, debugging, and software engineering benchmarks
• NiteAgent — AI agent development, frameworks, and production patterns

Cross-links automatically generated from None.