Kimi K2.7 Code on Fireworks: better agents, lower cost per task, available day-0

Today Moonshot released K2.7 Code, the latest in their K2 line of coding models, and we are launching Day-0 support for it on Fireworks. The architecture is the same one the K2 family has been using: 1T total parameters, 32B active per token, a 256K context window, tuned for long-horizon agentic coding.Pricing matches Moonshot’s public rates: $0.95 per 1M input tokens, $4.00 per 1M output tokens, and $0.19 per 1M on cache hits.
Pricing matches Moonshot’s public rates: $0.95 per 1M input tokens, $4.00 per 1M output tokens, and $0.19 per 1M on cache hits.

Model details: https://fireworks.ai/models/fireworks/kimi-k2p7-code

Model path: accounts/fireworks/models/kimi-k2p7-code

Here’s the bit we find interesting.

The number that caught our eye isn’t on the benchmark table. K2.7 Code uses roughly 30% fewer reasoning tokens than K2.6, and it still scores higher on coding evals:

• +21.8% on Kimi Code Bench v2
• +11.0% on Program Bench
• +31.5% on MLS Bench Lite

We think the ordering here is the story. “Spends fewer tokens thinking” is the headline, and “also scores higher” is the footnote. Usually it’s the other way around. If you’ve ever watched a coding agent talk itself in circles before doing the obvious thing, you already know why this matters more than another point or two on a benchmark.

Open in lightbox

A person reads and hears maybe 100,000 words a day, across everything. That’s around 130K tokens. If you talked to a frontier model continuously, 24 hours a day, no sleep, you’d still produce well under a million tokens of conversation. A few dollars of inference. For as long as humans were the only ones reading the output, the per-token price was almost a rounding error. We were the bottleneck, and we’re slow.

Agents changed that. A single long-horizon coding task fires off dozens of model calls, tool invocations and reasoning chains, and it can chew through more tokens in an hour than you’ll read in a year. Run a few of those in parallel across a team and you’re into billions of tokens a day without trying.

And the costs don’t add up, they compound. Each turn carries the history forward, so verbose reasoning from turn 3 gets re-read as input on every turn after. You pay to write those tokens once and to read them back dozens of times, plus the slower loops and retries that never show up on a pricing page.

So a 30% cut in reasoning tokens is, to us, worth more than a 30% cut in price. A price cut applies once, per token. Token discipline applies across the whole trajectory: shorter generations, smaller context on every following turn, faster loops, fewer retries.

Which leads to the conclusion we keep coming back to: per-token rate cards made sense when humans were the consumers. With agents, the real unit of spend is the completed task. K2.7 Code’s rate card is nearly identical to K2.6’s, but its cost per finished task is meaningfully lower. That’s the number we’re be watching.

Model efficiency and serving efficiency stack. Fewer reasoning tokens on fast inference means shorter loops, not just cheaper ones.

Kimi K2.7 Code runs on Fireworks serverless, with three serving options built for bursty agentic traffic:

• Standard is the elastic, pay-per-token default. Every existing API call works as-is.
• Priority is for traffic that cannot afford to be shed during peak congestion. Set service_tier: "priority" and your requests get stronger admission control on the shared fleet, at roughly 1.5x Standard pricing. Reliability, not speed, is the knob.
• Fast is a separate high-throughput serving path that runs the same weights at 100+ generated tokens per second, at roughly 2x Standard. You reach it through a dedicated Fast model id, not a per-request flag. For agent loops where wall-clock time is the constraint, Fast plus K2.7’s shorter reasoning chains is the compounding play. Fast for K2.7 Code is not (quite) ready, but coming shortly.

We understand that agentic traffic is bursty and uneven. That doesn’t mean devs should have to reserve GPUs or predict traffic shapes to get production-grade reliability. Route your critical agent traffic to Priority per request, and leave everything else on Standard. Fast adds a high-throughput path for latency-bound loops once it lands.

Kimi K2.7 Code is available now:

1234567
curl https://api.fireworks.ai/inference/v1/chat/completions \
  -H "Content-Type: application/json" \
  -H "Authorization: Bearer $FIREWORKS_API_KEY" \
  -d '{
    "model": "accounts/fireworks/models/kimi-k2p7-code",
    "messages": [{"role": "user", "content": "Refactor this function for readability."}]
  }'

Try it in the model playground, or read the serverless docs to configure Priority.