Frontier Models
MoonshotAI Releases Kimi K3 as Largest Open 2.8T MoE Frontier Model
The July 16 launch provides immediate API access at $3 per million input tokens while scheduling full open weights for July 27, positioning the model as a leader in specific coding benchmarks despite trailing overall proprietary leaders.
Kimi K3 is a 2.8 trillion parameter mixture-of-experts model released by MoonshotAI on July 16, 2026, that activates 16 of 896 experts and supports a 1 million token context window along with native multimodal inputs.
MoonshotAI introduced Kimi K3 as its flagship frontier model on July 16, 2026. The announcement highlighted the model's scale and efficiency features. API access became available immediately at the stated pricing tiers.
The company positions Kimi K3 for long-horizon coding and reasoning workloads. It builds on prior Kimi chatbot developments documented in public records. The release includes native vision and video capabilities.
What background preceded the Kimi K3 launch?
MoonshotAI has pursued large-scale model development for several years. The firm released earlier versions of the Kimi chatbot before advancing to this parameter class. Public documentation shows incremental improvements in context length and architecture.
Competitive pressure from closed models has driven open-weight strategies among several developers. Kimi K3 enters a market where parameter counts have scaled rapidly. The 2.8 trillion total reflects this trend toward larger mixture-of-experts systems.
The July 16 date aligns with a pattern of mid-year frontier releases. MoonshotAI chose immediate API availability to gather user feedback. Full weight release follows ten days later on July 27.
What new capabilities arrive with Kimi K3?
Kimi K3 introduces a 1 million token context window. This length exceeds many current commercial offerings. Native multimodal support extends input types to include vision and video.
The model tops the Frontend Code Arena at 1679 points. It ranks first in six of seven evaluated domains. Performance remains competitive with closed systems on agentic tasks.
API pricing undercuts several proprietary alternatives at $3 per million input tokens. Output tokens cost $15 per million. High cache-hit rates further reduce effective costs for repeated queries.
How does the Kimi Delta Attention architecture function?
Kimi Delta Attention combines hybrid linear attention with Attention Residuals. The design targets faster decoding speeds. Reported gains reach 6.3 times improvement over baseline attention mechanisms.
Stable LatentMoE governs expert selection during inference. Only 16 experts activate from the total pool of 896. This selective activation maintains performance while controlling compute demands.
The architecture supports both coding and knowledge work applications. MoonshotAI states the model targets frontier intelligence across extended reasoning chains. Details appear in the official announcement.
What pricing and availability details apply to Kimi K3?
API endpoints opened on the same day as the model announcement. Developers can access the model through the Kimi platform documentation. Pricing remains fixed at the published rates regardless of context length.
Open weights release occurs on July 27, 2026. This timeline allows ten days of closed API operation. The move aligns with broader industry shifts toward reproducible research.
High cache-hit rates are expected to lower real-world costs. The pricing model favors long-context applications. Users benefit from the combination of scale and accessibility.
How does Kimi K3 compare to Claude Fable 5 and GPT-5.6 Sol?
Kimi K3 trails the top closed models on aggregate leaderboards. It placed third on the Artificial Analysis AI leaderboard upon release. Claude Fable 5 and GPT-5.6 Sol occupy the first two positions.
The open model leads on the specific frontend web development benchmark. Arena.ai data shows the 1679 point score ahead of competitors in that domain. Coding and agentic tasks represent its strongest areas.
Parameter count and context window provide direct comparison points. The 2.8 trillion total and 1 million token support differentiate Kimi K3 from smaller open alternatives. Open weights add a reproducibility advantage.
| Feature | Kimi K3 Specification |
|---|---|
| Total Parameters | 2.8 trillion |
| Active Experts | 16 out of 896 |
| Context Window | 1 million tokens |
| Architecture | Kimi Delta Attention with Attention Residuals |
| API Input Price | $3 per million tokens |
| Open Weights Date | July 27, 2026 |
What market and stakeholder implications follow from the release?
Open weights availability within ten days lowers barriers for downstream developers. Researchers gain the ability to inspect and fine-tune the full model. This transparency contrasts with closed API-only systems.
Pricing pressure may influence future commercial offerings. The $3 input rate challenges higher-tier proprietary models. Agentic and coding use cases stand to benefit most from the combination of scale and cost.
Stakeholders in the open-source community receive a new reference point. The 2.8 trillion parameter class sets a benchmark for subsequent releases. MoonshotAI's timeline signals continued investment in accessible frontier models.
What expert reactions have emerged around Kimi K3?
Early commentary focuses on the benchmark leadership in web development tasks. Observers note the rapid shift from API-only to open weights. The architecture details have drawn attention from efficiency researchers.
Some analysts highlight the gap between overall leaderboard position and domain-specific strength. The model demonstrates that open systems can compete on narrow capabilities. Pricing transparency adds another dimension to the discussion.
The ten-day window to open weights has prompted questions about deployment readiness. Developers are preparing evaluation suites ahead of the July 27 date. Community forums show active planning for fine-tuning experiments.
What comes next after the Kimi K3 launch?
MoonshotAI has indicated further iterations on the Kimi series. The current release serves as a foundation for community contributions once weights are public. Additional multimodal enhancements remain possible.
The open weights phase will enable independent verification of the reported 6.3 times decoding speedup. Researchers will test long-context performance at scale. Integration into agent frameworks is expected to follow quickly.
Market observers will track adoption metrics after the July 27 release. Pricing adjustments or new tiers could emerge based on usage patterns. The model establishes a new baseline for open frontier systems.
- API access opened on July 16, 2026.
- Full weights scheduled for July 27, 2026.
- Community fine-tuning expected immediately after weight release.
- Further Kimi series updates anticipated in subsequent quarters.
Today, we are introducing Kimi K3 — our most capable model. Kimi K3 is a 2.8T-parameter model built on our Kimi Delta Attention and Attention Residuals, with native vision capabilities and a 1-million-token context window. It is the world's first open 3T-class model, designed for frontier intelligence across long-horizon coding, knowledge work, and reasoning.Moonshot AI
Frequently asked
What context window does Kimi K3 support?
Kimi K3 supports a 1 million token context window. This length enables processing of extensive documents and codebases in a single pass. The capability pairs with the model's multimodal inputs for complex tasks.
When will the open weights become available?
The full model weights are scheduled for release on July 27, 2026. This follows the July 16 API launch by ten days. Developers can prepare evaluation and fine-tuning pipelines in advance.
How does Kimi K3 pricing compare to competitors?
Input tokens cost $3 per million and output tokens cost $15 per million. The rates include support for high cache-hit scenarios. This structure positions the model competitively against several proprietary offerings.