Moonshot AI has released Kimi K2.8 Preview, a new mid-tier model that sits between its coding-focused K2.7 Code and flagship K3, adding multimodal capabilities and expanded context length while keeping costs lower than the flagship. The model went live on September 11, 2026, inside Kimi Code and Kimi Work with no configuration changes required for existing users.

What Changed in Kimi K2.8 Preview?

The most significant upgrade is context length. Every Kimi Code membership tier now gets a 1 million token context window, a substantial jump from the 262,144 tokens available on K2.7 Code. To put that in perspective, 1 million tokens can process roughly 750,000 words at once, making it possible to feed an entire codebase, lengthy document, or hours of transcript into a single prompt without breaking it into smaller chunks.

Kimi K2.8 Preview also introduces multimodal input, accepting text, images, and video alongside text-only output. This represents a departure for Moonshot, which released K2.7 Code and K3 as open-weight models; K2.8 Preview is closed and available only through Moonshot's own API and apps.

Reasoning control is another meaningful addition. The model supports three thinking effort levels, low, high, and max, matching the structure Moonshot introduced with K3. It runs at max effort by default, and Moonshot reports that thinking efficiency has improved significantly compared with K2.7 Code, alongside broader gains in coding and agent tasks.

How Does K2.8 Preview Compare to Other Moonshot Models?

Moonshot positions K2.8 Preview firmly in the middle of its lineup, more capable than K2.7 Code but not a replacement for K3. For context, K3 is Moonshot's 2.8 trillion parameter open-weight flagship, released in July 2026 with its own 1 million token context window and a mixture-of-experts architecture built for long-horizon coding and reasoning work. Moonshot claims K2.8 Preview performance is close to K3, but independent benchmarks confirming this have not yet been published.

In the broader Chinese AI landscape, Moonshot's Kimi models rank among the top performers. Kimi K3 scored 1543 on the AA-Briefcase agent benchmark, beating GPT-5.6 Sol (1501) and trailing only Claude Fable 5 (1574), making it the current agent king for autonomous tasks like coding and data analysis. Kimi K2.6, an earlier model, achieved 80.2% on SWE-Bench Verified, a software engineering benchmark, beating GPT-5.4 and competing with Claude Opus at a fraction of the price.

How to Choose the Right Kimi Model for Your Needs

• For Budget-Conscious Teams: Kimi K2 remains the cheapest option in the family at $0.60 per million input tokens and $2.50 per million output tokens, making it suitable for teams that need decent performance without premium pricing.
• For Coding-Heavy Workloads: Kimi K2.6 is the SWE-Bench king with 80.2% accuracy on software engineering tasks, offering open-weight availability and strong performance on real-world bug fixing and feature implementation.
• For Long-Context Agent Work: Kimi K2.8 Preview bridges the gap with its 1 million token context window and multimodal input at mid-tier pricing, ideal for teams needing vision and language capability without K3's cost.
• For Maximum Performance: Kimi K3 remains the flagship at $3 input and $15 output per million tokens, delivering 2.8 trillion parameters and the highest benchmark scores for autonomous agent tasks.

Third-party API gateway TokenRa lists K2.8 Preview pricing at roughly $0.80 per million input tokens and $3.35 per million output tokens, with cached read tokens at $0.14 per million. Moonshot has not published first-party API pricing for the preview model as of this writing, so gateway pricing should be treated as an indicator rather than an official rate card.

Why Release a Mid-Tier Model Right Now?

The timing aligns with Moonshot's push to widen usage without running every request through its most expensive model. Moonshot has seen rapid revenue growth since K3 shipped and has reportedly begun a Hong Kong IPO process, framing K2.8 Preview as a way to lower the cost of giving more users long-context and agentic features. Handing every membership tier a 1 million token window on a cheaper model, rather than reserving that context length for flagship-only plans, fits a company trying to scale usage while keeping inference costs predictable.

This mirrors a pattern other AI labs have followed in 2026: ship a flagship, then follow with a lighter, faster sibling that keeps most of the capability at a lower price. The trade-off between raw capability and everyday cost shows up across the industry, as companies balance performance with accessibility.

The "preview" label is doing real work here. Moonshot is treating K2.8 Preview as an early-access checkpoint, not a finished product, and the company is gathering feedback before deciding how the model fits into its permanent lineup. Until third-party evaluators publish scores on standard suites like LiveCodeBench or SWE-Bench, the comparison between K2.8 Preview and K3 remains Moonshot's word against its own flagship.

For developers and teams evaluating Moonshot's lineup, K2.8 Preview represents a meaningful middle ground, offering expanded capabilities at a cost point between the budget K2 line and the premium K3 flagship. The real value will come down to cost, latency, and reliability across actual coding workloads rather than marketing claims alone.