Kimi K2.7 Code.
Kimi K2.7 Code is a modified-MIT coding agent model, 1T total and 32B active, with image and text input and a 256K context, built on Kimi K2.6.
8× Nvidia A100 SXM4 · $5.73/hr.
Most-aggressive quantisation we have a working recommendation for. Lower precision = less VRAM = cheaper hardware, at a small accuracy cost.
Cheapest hosted endpoints.
Variants in the Kimi family.
Moonshot's frontier open-weight MoE — 1T total, 32B active.
Moonshot's open-weight reasoning variant — extended chain-of-thought training...
Multimodal agentic variant — adds a vision encoder to the K2 backbone.
Long-horizon coding + autonomous-execution upgrade over K2.5.
Kimi K3 is Moonshot AI's open-weight multimodal agentic model, 2.8T parameter...
Kimi K3.1 Pro is Moonshot AI's flagship open-weight multimodal reasoning mode...
Kimi K2 0905 is a modified-MIT mixture-of-experts model, 1T total and 32B act...
Cheapest hardware per quantisation.
Each row is one quantisation tier (the same weights compressed differently). Lower precision → lower VRAM → cheaper hardware, at the cost of small accuracy loss. $/小时为当前最低报价。多卡配置在有同规格实例出租时使用真实实例价格,否则为单卡价格乘以卡数。
| Quantisation | Cheapest GPU config | Total VRAM | Live $/hr | tokens/sec | |
|---|---|---|---|---|---|
|
FP8
FP8 — 8-bit float (Hopper / Blackwell)
|
1536 GB |
$22.60/hr
Cyfuture AI · 按需
|
— | Compare → | |
|
INT4
INT4 — 4-bit integer (~4× VRAM saving)
|
640 GB |
$5.73/hr
估算:8 × 单卡价格
Nosana · 按需
|
— | Compare → |
What it costs per month across providers.
Estimate your monthly bill for Kimi K2.7 Code across every host that publishes per-token pricing. Slide your token volumes; the chart + table re-rank cheapest-first.
最便宜的服务商在左侧。
每月总成本 — 输入与输出 token 合计。
账单明细。
租 GPU,而不是按 token 付费。
对于 Kimi K2.7 Code 这类开放权重模型,你可以租一块 GPU 自己提供推理服务。算法:最便宜的 GPU 租价 × 每月 730 小时。电费已包含在租用的每小时价格中。
假设 GPU 全天 24 小时运行,利用率约 85%。如果流量时高时低,API 会更省钱,GPU 则可能更贵(空闲时段仍要付租金)。盈亏平衡分析见自托管与 API 盈亏平衡工具。
About Kimi K2.7 Code.
Kimi K2.7 Code is a coding-focused agentic model from Moonshot AI, built on Kimi K2.6, with open weights under a modified MIT licence. It is a mixture-of-experts model with 1T parameters in total and 32B active, plus a 400M vision encoder, so it takes images as well as text. Context length is 256K tokens. Moonshot aims it at long software engineering tasks carried through to the end, and says it spends fewer thinking tokens than K2.6.