A 1,000,000-Token Context Window Across All Tiers

Every Kimi Code membership tier now includes a 1,000,000-token context window, up from 262,144 on K2.7 Code — roughly a fourfold expansion. For teams that previously had to split large inputs across multiple calls and stitch results together, this collapses the workflow into a single pass. It matters most where coherence across a long input is the point, not just retrieval.

Business outcome: eliminates chunking and retrieval overhead for long-document and large-codebase analysis.

Text, Image, and Video Input in One Model

K2.8 Preview accepts text, image, and video input natively, rather than routing different media types through separate models. For media, training, and content operations teams, that simplifies pipeline architecture — one endpoint handles mixed-media analysis. It also reduces the number of vendor relationships a team needs to maintain for multimodal work.

Business outcome: consolidates multimodal analysis into a single integration instead of multiple specialised models.

Adjustable Thinking-Effort Levels Aligned With K3

The model supports adjustable thinking-effort levels that mirror K3's reasoning tiers, letting teams dial reasoning depth up or down per task. High-stakes analysis can run at deeper reasoning; routine extraction can run fast and cheap. This is a cost-control lever as much as a quality lever, since reasoning depth typically drives token consumption.

Business outcome: lets teams match compute spend to task criticality rather than over-provisioning every request.

Same API Model ID as K2.7 Code — Zero Config Changes

K2.8 Preview retains the same API model ID as K2.7 Code, so switching requires no configuration changes on the client side. For teams already running Kimi in production, that removes the migration project that usually accompanies a model upgrade. It is a small detail that carries outsized operational value.

Business outcome: upgrades existing Kimi deployments without engineering time spent on reconfiguration or regression testing.

Mid-Tier Slot Between K2.7 Code and K3

The model is deliberately positioned between Kimi K2.7 Code and the flagship Kimi K3, giving buyers a middle option rather than a binary choice between coding-specialised and frontier-tier. For teams that need more than K2.7 Code offers but cannot justify K3 economics, that middle slot is the point. It also gives Moonshot a tier to move price-sensitive customers into without losing them.

Business outcome: provides a cost-appropriate tier for teams that have outgrown the coding model but do not need flagship pricing.

Bundled Into Every Kimi Code Membership Tier

The expanded context window is available across every Kimi Code membership tier rather than gated behind the top plan. That means smaller teams get the same context ceiling as enterprise subscribers, which is unusual in a market where context length is often a premium upsell. Buyers evaluating tiers should confirm current plan structures directly, since published pricing was not available in the material reviewed.

Business outcome: removes the tier-gating that typically forces smaller teams to pay enterprise rates for long-context work.

• Context window is genuinely large and not tier-gatedA 1,000,000-token window available on every Kimi Code tier is unusual — most vendors reserve long context for premium plans.
• Zero-friction migration from K2.7 CodeThe unchanged API model ID means existing deployments can adopt K2.8 Preview without touching client configuration.
• Multimodal input in a single modelText, image, and video support in one endpoint reduces the number of integrations a team has to maintain.
• Adjustable reasoning depth gives cost controlThinking-effort levels aligned with K3 let teams match spend to task criticality instead of over-provisioning.

• No independently verified benchmark scoresOutside sources note that no third-party-verified benchmark results exist yet for K2.8 Preview, so capability claims rest on vendor positioning alone.
• The 'close to K3' claim is qualitative onlyMoonshot's own performance statement is qualitative — 'overall capability close to K3' — with no published numbers behind it.
• Pricing is not publicly listedWithout published tier pricing, buyers cannot compare total cost of ownership against alternatives without contacting the vendor directly.
• Professional RealityTreat the close-to-K3 positioning as unproven until independent evaluations publish. If your decision hinges on frontier-tier reasoning quality, wait for third-party benchmarks or run your own evaluation on representative tasks before migrating production workloads.

Whole-Codebase Analysis for Engineering Teams

Development teams can load large repositories or multi-module systems into a single 1M-token session, avoiding the retrieval scaffolding that fragments code understanding. Combined with the unchanged API model ID, existing Kimi AI integrations can adopt this without reconfiguration. The result is fewer moving parts in the analysis pipeline.

Long-Document Review in Legal and Financial Work

Contracts, filings, and research reports that previously had to be chunked can be processed in one pass, preserving cross-references and coherence. For teams in regulated industries, that reduces the risk of context loss between chunks. It also simplifies audit trails, since a single session covers the whole document.

Video Content Analysis Without Pre-Transcription

Native video input means media and training teams can analyse footage directly rather than running a separate transcription step first. That shortens the pipeline and removes a failure point. For content operations handling large video libraries, the consolidated workflow is the main draw.

Cost-Tiered Reasoning for Mixed Workloads

Teams running a mix of routine extraction and deep analysis can dial thinking effort per task, spending compute where it changes the outcome. This is particularly useful for operations teams processing high volumes of low-complexity requests alongside occasional complex ones. It turns reasoning depth into a budget lever rather than a fixed cost.

Confirm your current Kimi Code membership tier and verify the 1,000,000-token context window is included on your plan via the official pricing page.

If you already run K2.7 Code in production, test K2.8 Preview against your existing API model ID — no configuration changes should be required.

Run a representative long-context task (a full codebase module, a long contract, or a video file) and compare output quality against your current model.

Establish thinking-effort defaults per workload type, so routine tasks run fast and high-stakes analysis runs at deeper reasoning.