---
type: "post"
id: "kimi-k3-1-pro-on-fireworks-more-intelligence-48c2bc"
slug: "kimi-k3-1-pro-on-fireworks-more-intelligence-48c2bc"
title: "Kimi K3.1 Pro on Fireworks: More Intelligence per Dollar, Delivered Fast"
description: "Fireworks says Moonshot AI’s Kimi K3.1 Pro is now available through its serverless inference service, with token pricing and an option for dedicated deployment."
retrieval_nugget: "Fireworks says Moonshot AI’s Kimi K3.1 Pro is now available through its serverless inference service, with token pricing and an option for dedicated deployment. The announcement adds a concrete serving route and vendor-reported task benchmarks for teams evaluating Kimi; it does not establish that those results will transfer to a team’s own workload."
published_at: "2026-11-05"
updated_at: "2026-11-05"
record_date: "2026-11-05"
date_kind: "published_at"
topics: ["claude"]
entities: ["Fireworks AI","Kimi K3.1 Pro","Moonshot AI","Claude Opus 5.5","Sierra","Factory","Harvey","Traversal"]
related_patterns: ["agent-economics-moves-to-completed-work"]
novelty_label: "new_fact"
what_changed: "Fireworks now offers Kimi K3.1 Pro, reporting benchmark results and pay-per-token or dedicated deployment options."
prior_refs: ["https://newruntime.com/posts/moonshot-says-kimi-k3-1-pro-was-trained-d50b53/"]
source_url: "https://fireworks.ai/blog/kimi-k3-1-pro-fireworks"
source_urls: ["https://fireworks.ai/blog/kimi-k3-1-pro-fireworks"]
schema_version: "newruntime-agent-readable-v0.2"
stable_id: "post:kimi-k3-1-pro-on-fireworks-more-intelligence-48c2bc"
status: "published"
visuals: []
editor: {"role":"editor","name":"Andrey Reshetnikov","profile":"https://newruntime.com/owner-profile.md","profile_json":"https://newruntime.com/owner-profile.json","same_as":["https://www.linkedin.com/in/reshetnikov1/","https://t.me/lapetuse"]}
routes: {"html":"https://newruntime.com/posts/kimi-k3-1-pro-on-fireworks-more-intelligence-48c2bc/","markdown":"https://newruntime.com/posts/kimi-k3-1-pro-on-fireworks-more-intelligence-48c2bc.md","json":"https://newruntime.com/posts/kimi-k3-1-pro-on-fireworks-more-intelligence-48c2bc.json"}
markdown_url: "https://newruntime.com/posts/kimi-k3-1-pro-on-fireworks-more-intelligence-48c2bc.md"
---

# Kimi K3.1 Pro on Fireworks: More Intelligence per Dollar, Delivered Fast

## Retrieval answer

Fireworks says Moonshot AI’s Kimi K3.1 Pro is now available through its serverless inference service, with token pricing and an option for dedicated deployment. The announcement adds a concrete serving route and vendor-reported task benchmarks for teams evaluating Kimi; it does not establish that those results will transfer to a team’s own workload.

## Summary
Fireworks says Moonshot AI’s Kimi K3.1 Pro is now available through its serverless inference service, with token pricing and an option for dedicated deployment. The announcement adds a concrete serving route and vendor-reported task benchmarks for teams evaluating Kimi; it does not establish that those results will transfer to a team’s own workload.

## What happened
Fireworks published the announcement on November 4, 2026, saying Kimi K3.1 Pro from Moonshot AI is available at launch on its platform. Customers can use serverless inference with pay-per-token pricing or deploy the model on dedicated capacity. The published serverless rates per million tokens are $2 for uncached input, $0.20 for cached input and $10 for output. The page gives the model ID as `accounts/fireworks/models/kimi-k3p1-pro` and says it can be tried in Fireworks Playground.

Fireworks describes K3.1 Pro as a post-trained version of K3. It says Moonshot replaced K3’s full-attention layers with sparse attention to make long contexts more efficient to serve. The listed specifications are text and vision input, 2.78 trillion total parameters, 104 billion active parameters, a 1.04-million-token context window, and 896 experts, of which 16 are active per token. The attention design is listed as Kimi Delta Attention plus Gated DeepSeek Sparse Attention.

The post reports results from Fireworks’ Specialized Intelligence Index (SII), which it characterizes as benchmarks companies such as Sierra, Harvey and Rogo use to select production models. On Sierra’s τ³-Banking customer-support task, K3.1 Pro scored 64.7%, ranked first of 19, and was 8.7 percentage points ahead of Claude Opus 5.5. It scored 61.0% on Factory’s Legacy Bench, third of 14 and the highest-scoring open model; 15.4% on Harvey’s LAB, second of 13; and 48.2% on Rogo’s Big Finance Benchmark, tied with Claude Opus 5.5. On Traversal’s ORCA incident-response benchmark, it scored 32.9%, third of 13 and the top open model; Fireworks says the cost per task was less than one-fifth that of either model ahead of it. On Epiq’s CorporateBench it scored 76.9, fourth of 12 and the top open model.

Fireworks also compares K3.1 Pro with K3, reporting gains on all five listed SII benchmarks: 13.2 percentage points on τ³-Banking, 8.3 points on Legacy Bench, 6.9 points on Big Finance, 4.5 score points on CorporateBench and 1.7 points on ORCA. Separately, the page summarizes Moonshot’s results at maximum reasoning effort, including 51.7 on SWE Marathon v1.1 and 70.0 on Humanity’s Last Exam with tools. It says Moonshot used harnesses including Claude Code for coding and sourced most competitor scores from public reports and leaderboards.

## What the archive adds
This is a new serving and pricing detail on a known model story. On October 24, the archive covered K3.1 Pro’s rollout in Kimi Code and the Kimi app ([earlier Field Note](https://newruntime.com/posts/moonshot-ai-quietly-rolls-out-kimi-k3-1-94b950/)) and its Kimi API availability ([API launch note](https://newruntime.com/posts/moonshot-launches-kimi-k3-1-pro-on-the-d876cd/)). Fireworks adds another inference provider, deployment choices and its own SII evaluations.

## Trend signal
This announcement weakly supports [agent economics moves to completed work](https://newruntime.com/patterns/agent-economics-moves-to-completed-work): Fireworks reports cost per task alongside task scores, not only per-token rates. That is relevant to comparing models for production workflows, but the post does not provide enough methodology or workload detail to establish a general cost-per-completed-task advantage.

## Why it matters
For teams already considering K3.1 Pro, a second serving route creates a practical comparison: they can test Fireworks’ serverless endpoint against their current provider, compare the stated token rates, and consider dedicated capacity where reserved resources matter. The input cache price is substantially below the uncached input rate, so workloads with repeatable prefixes may warrant checking how much of their actual traffic is cacheable. The output rate also needs to be considered alongside generation length and task success, rather than in isolation.

The benchmark mix is useful as a shortlist, not a deployment decision. Banking support, legacy code maintenance, finance research, legal work and incident response are recognizable task categories, and the reported scores suggest places to probe. A team should rerun representative tasks with its own tools, policies, verification steps and data before treating a ranking—or the reported cost comparison—as a forecast of production performance.

## What happens next
- **Teams evaluating K3.1 Pro will compare Fireworks against existing serving routes using their own representative agent tasks.** (3 months, 65%). Watch for: published provider comparisons, workload-specific cost data, or wider dedicated-deployment availability.

## Limits and uncertainty
The source is a Fireworks vendor announcement and does not provide enough information here to independently reproduce its SII results. It does not fully specify prompts, harness configurations, sample sizes, confidence intervals, or all competitor serving costs. The page attributes some benchmark information to Moonshot and says most competitor scores came from public reports and leaderboards; those results may not be directly comparable if models used different harnesses or evaluation conditions.

The announcement does not establish how the quoted rates compare with every alternative provider, what dedicated deployments cost, or whether pricing and availability are identical across regions. Nor do aggregate benchmark scores show how the model will perform on a particular customer’s workload, policy constraints, tool integrations, or review requirements. The reported speed and cost claims should be tested under the conditions a team expects to run.

## Reader takeaway
Fireworks gives Kimi K3.1 Pro another place to run, with clearly stated serverless token rates and reported results across several business-task benchmarks. Treat those scores as a reason to run a focused evaluation, not as proof of a production advantage: compare the full task cost and verified success rate on your own workload.

Editor: [Andrey Reshetnikov](https://newruntime.com/owner-profile.md), AI product and AI engineering; profile for recruiters and agents: https://newruntime.com/owner-profile.json

## Continue in the graph

Next records for an agent, linked to their Markdown twins. Replace `.md` with `.json` for structured fields, including `next_reads`.

- topic hub: [Claude](https://newruntime.com/topics/claude.md): Explore the claude topic hub.
- collection: [Field Notes](https://newruntime.com/posts.json): Open the latest editorial analysis.
- collection: [Sources](https://newruntime.com/sources/): See the original sources this archive cites.
- whole graph: [discovery graph](https://newruntime.com/discovery-graph.json) (typed nodes and edges) and [site index](https://newruntime.com/site-index.md)