Moonshot adds HighSpeed mode to Kimi K2.7-Code, boosting output by 6x

If you have ever used a reasoning-heavy coding model, you know the frustration: you ask for a minor fix, and the cursor hangs for twenty seconds while the model writes a philosophical essay to itself before spitting out two lines of syntax.

Moonshot AI is trying to cut through that latency. Just days after releasing Kimi K2.7-Code on June 12, the Beijing lab rolled out HighSpeed mode on June 15 for its Kimi Code Beta, delivering an immediate 6x throughput boost.

The new mode pushes around 180 tokens per second on median coding tasks and hits up to 260 tokens per second on shorter inputs. For teams running autonomous agent loops that process hundreds of parallel requests, that speed bump changes everything.

Trimming the reasoning bloat

Under the hood, the architecture has not changed at all compared to K2.5 and K2.6.

It is still a trillion-parameter Mixture-of-Experts setup split across 384 specialist sub-networks, activating eight specialists and one shared expert per token to keep active compute around 32 billion parameters. It also retains the DeepSeek-pioneered Multi-head Latent Attention to fit a 256K context window without frying memory banks.

Crucially, K2.7-Code cannot skip its internal monologue. It operates strictly in thinking mode, carrying reasoning states across turns. But Moonshot claims the new model burns 30% fewer reasoning tokens on code tasks.

Instead of revamping the model layers, Moonshot simply adjusted its post-training mix. By heavily weighting coding runs, they forced the model to generate tighter, shorter reasoning paths. It takes aim at the well-known overthinking problem, where excessive internal tokens burn compute without adding real accuracy.

Fast access, zero independent benchmarks

Moonshot made sure developers can get their hands on it immediately. Weights are up on Hugging Face under a Modified MIT License, API access costs $0.95 per million input tokens and $4.00 per million output tokens, and its terminal agent Kimi Code runs $19 a month.

Yet there is a massive catch: Moonshot has not submitted K2.7-Code to a single independent benchmark.

The claims of 30% token reduction and cutting-edge performance rest solely on five internal tests created and run by Moonshot itself. For now, anyone pulling the model has to take the lab at its word.

Why it matters

We think this release exposes the double-edged sword of current reasoning models.

For engineers, long-winded reasoning chains translate directly to wasted cloud credits and sluggish feedback loops. A 6x speed jump and fewer wasted tokens make agentic coding far more practical. With open weights and a cheap API entry point, testing it against your own test suite is basically a no-brainer.

It also highlights Moonshot's blistering release cadence. Since dropping the original K2 in July 2025, this is their fifth major iteration. Backed by a $20 billion valuation and cash from Alibaba, Tencent, and China Mobile, the lab is shipping updates faster than most closed-source giants.

Still, skipping independent evaluation looks messy. Slashing reasoning length can easily backfire into underthinking when edge cases pop up. Until community evals put real numbers behind the marketing claims, production teams should probably think twice before replacing battle-tested backends.