Breaking

Moonshot Rolls Out HighSpeed Variant of Its Kimi K2.7 Code Model

June 15, 2026 at 06:42 EDT

Foundation Models
Software Dev & Coding
Open Source

Moonshot AI on June 15, 2026 made "Kimi K2.7 Code HighSpeed" (model ID: kimi-k2.7-code-highspeed), a faster variant of its open-weight coding-focused model Kimi K2.7 Code, available on the Kimi API platform. While identical to the standard model, the variant pushes inference speed up to six times higher, and is being rolled out in phases to Beta Program members, developers and Business users for now.

June 15, 2026 · Moonshot AI

Kimi K2.7 Code HighSpeed: up to 6× faster coding, 30% less overthinking

An accelerated inference mode for Moonshot's open-weight, agentic coding model — pitched as a low-cost alternative to Claude Opus and GPT-5.5, now in limited rollout.

6×

Generation speed vs prior version

~180

Tokens/sec, median coding tasks (up to 260 on short contexts)

−30%

Reasoning tokens used (curbing overthinking)

Total params (MoE, 32B active · 256K context)

Benchmark gains — K2.6 vs K2.7 Code

Company-reported, thinking mode enabled. Each pair: left column = K2.6, right = K2.7 Code.

50.9

62.0

Kimi Code Bench v2

+21.8%

48.3

53.6

Program Bench

+11.0%

26.7

35.1

MLS Bench Lite

+31.5%

API pricing — per million tokens

$0.95

Input

$4.00

Output

$0.19

Cache hit

HighSpeed surcharge undisclosed · open weights under Modified MIT · thinking-mode only.

Practitioners like

"A better default for longer coding jobs"
"Less overthinking means fewer stalls and lower cost"
Stable on long-horizon agentic tasks: repo navigation, multi-step debugging

Caution flags

Benchmarks lean on in-house suites; no SWE-Bench Verified or KernelBench validation
Some independent tests show regressions vs K2.6 — "more honest but not more capable"
HighSpeed still in limited rollout — speed claims unverified at scale

Continue reading

The rest of this article is for AI News Blitz readers. Choose an option below to keep reading.