K2.8 Preview is the model to reach for if you already use Kimi Code and want K3-level coding help without a premium plan, since it now defaults to three thinking-effort levels instead of six. It has no independent benchmark yet, so treat any published score for it as unverified.
Kimi K2.8 Preview is Moonshot AI's mid-tier coding model, released in September 2026 with a 1,048,576-token context window opened to every Kimi Code membership tier. It now serves the same kimi-for-coding model ID that Moonshot's prior coding release used, with performance the vendor says approaches flagship Kimi K3.
Provider: Moonshot AI · Family: Kimi
More about Moonshot AI on HokAI
Context window: 1,048,576 tokens
Input modalities: text, image · Output: text, tool-calls, code
About Kimi K2.8 Preview
Kimi K2.8 Preview is a mid-tier coding and agentic model from Moonshot AI, the Chinese AI lab behind the Kimi line. Moonshot rolled it out inside the Kimi Code product on September 11, 2026, serving it at the same kimi-for-coding API model ID that previously pointed to Kimi K2.7 Code, so existing clients and third-party integrations needed no configuration change to receive the upgrade. It sits between K2.7 Code (released June 12, 2026 with a 256K context window) and the flagship Kimi K3 (July 16, 2026, 2.8 trillion parameters) in Moonshot's lineup, and Moonshot describes its overall performance as close to K3 with more efficient reasoning.
The clearest confirmed upgrade is context: K2.8 Preview opens Kimi Code's full 1,048,576-token window to every membership tier, a limit Moonshot had previously reserved for its higher plans. It exposes three adjustable thinking-effort levels, low, high, and max (max is the default), a narrower set than K3's six-level scheme of none, low, medium, high, xhigh, and max. Moonshot's own Kimi Code changelog confirms both facts directly. That changelog entry does not itemize input modalities for this release, so this record follows K3's own verified capability set (text and image input, with no confirmed audio or video input) rather than repeating the video-input claims some third-party trackers have made for K2.8 without a Moonshot source.
Moonshot has not published a model card, parameter count, or architecture description specific to K2.8 Preview. Because it is served at the same model ID as K2.7 Code with, in Moonshot's own words, no configuration change required, it most likely shares that model's Mixture-of-Experts transformer design (1 trillion total parameters, 32 billion active per token), but that is an inference from the shared endpoint rather than a confirmed K2.8 specification, and should be treated as unverified until Moonshot publishes one.
No independently verified benchmark score exists yet for K2.8 Preview. Moonshot's only public performance claim for it is qualitative: overall capability close to K3. As a verified reference point, K3 itself scored 93.5% on GPQA Diamond and 67.5% on DeepSWE in its own July 2026 release, ahead of GPT-5 and open peers like DeepSeek V4 and GLM-5.2 on several evals, but those figures belong to K3, not K2.8, and should not be assumed to carry over.
K2.8 Preview is not billed per token. It ships bundled inside Kimi's membership plans through the Kimi Code and Kimi Work product surfaces, the same way K2.7 Code was, rather than as a line item on Moonshot's developer API pricing page, which as of September 15, 2026 lists kimi-k3, kimi-k2.7-code, kimi-k2.7-code-highspeed, and kimi-k2.6 but no separate K2.8 entry. Kimi Code's own documentation names membership tiers including Moderato and Allegretto, but the specific monthly prices render client-side on Moonshot's pricing page and could not be independently confirmed for this record.
K2.8 Preview fits teams already inside the Kimi Code or Kimi Work ecosystem who want a faster, more efficient model for everyday coding and agent tasks without paying for a K3-tier plan, and who can tolerate a model that changes under the hood without a version bump, exactly what happened on September 11. It is the wrong choice for teams that need a documented, independently benchmarked model before adoption, a model reachable through the general developer API alongside tools like Claude Code, or committed open-weights access. The open-weighted Kimi K2.7 Code or the flagship K3 remain the better-documented options until Moonshot publishes K2.8's own model card.
Pricing
Not billed per token. K2.8 Preview rides on whatever Kimi membership plan is already tied to the account rather than a separate per-token charge. Moonshot's platform pricing page lists rates for its other current models but carries no separate K2.8 line item as of publication, and the exact monthly tier prices were not independently verifiable in a form this record could confirm.
Key Features
- 1M-token context on every tier: Kimi Code's full 1M-token context window is now available to every membership tier, a limit Moonshot previously reserved for its higher plans.
- Three-level adjustable reasoning: Supports low, high, and max thinking-effort levels (max by default), controllable per request in Kimi Code with the --effort flag.
- Drop-in upgrade at the same model ID: Served at the existing kimi-for-coding model ID, so nothing changed on the client side when Moonshot swapped the model answering that route.
- Bundled into existing membership: Not billed per token; included in your existing Kimi subscription rather than sold as a new premium add-on.
Pros
- Opens the full 1M-token context window to every Kimi Code membership tier instead of gating it to premium plans.
- Ships as a no-cost, no-config upgrade for anyone already calling the kimi-for-coding model ID.
- Moonshot reports meaningfully more efficient reasoning than the model it replaced at the same endpoint.
Cons
- No SWE-bench, GPQA, or other independently verified benchmark score has been published for this checkpoint.
- Not available through Moonshot's general developer API or pricing page, only inside Kimi's own products.
- No model card, parameter count, or system card exists yet, so architecture and safety behavior are inferred, not confirmed.
Frequently Asked Questions
How much does Kimi K2.8 Preview cost?
K2.8 Preview has no per-token price of its own; access comes from whichever Kimi subscription the account already holds. Moonshot's platform pricing page does not list a separate rate for it, and the specific tier prices render client-side on kimi.com in a way that resisted independent verification.
How does Kimi K2.8 Preview perform on benchmarks compared to Kimi K3?
No published SWE-bench, GPQA, or other quantified score exists for this checkpoint; the vendor's only claim is that overall capability now approaches flagship K3. For reference, the K3 release reported a GPQA Diamond result of 93.5 percent and a DeepSWE result of 67.5 percent, but until Moonshot verifies a number for K2.8 itself, treat any benchmark quoted for it as unconfirmed rather than equal to K3's.
Is Kimi K2.8 Preview open source?
No. Unlike Kimi K2.6 and K2.7 Code, which Moonshot released as open weights on Hugging Face under a Modified MIT license, K2.8 Preview has no Hugging Face repository or public license as of September 15, 2026. It is reachable only inside Moonshot's own Kimi Code and Kimi Work products at the kimi-for-coding model ID.
Does Kimi K2.8 Preview train on user data?
Moonshot has not published a data-retention or training policy specific to this checkpoint. No system card or privacy document exists yet for K2.8 Preview, so its retention and training-on-inputs behavior should be confirmed directly with Moonshot before sending sensitive code through it.
Who should use Kimi K2.8 Preview, and who should avoid it?
It suits teams already paying for a Kimi membership who want faster coding help than K2.7 Code without upgrading to a K3-tier plan. Teams needing a benchmarked, independently documented model or direct per-token API billing are better served today by K3 or the open-weighted K2.7 Code.