TokenMix Research Lab · 2026-07-17

Kimi K3 Review 2026: 2.8T Params, $3/$15 API, Benchmarks

Last Updated: 2026-07-17 Author: TokenMix Research Lab Data verified: 2026-07-17 - Kimi K3 technical blog, Kimi API pricing and quickstart documentation, live TokenMix model catalog, Artificial Analysis, Axios, and Xinhua

Kimi K3 is live with 2.8T parameters, 1M context, and a $3/$15 API. The launch is substantial, but not a clean victory: only max reasoning is available, the weights are promised by July 27, and independent testing finds high intelligence alongside slower-than-average, unusually verbose output.

Moonshot AI released Kimi K3 on July 16, 2026 across Kimi, Kimi Work, Kimi Code, and the official API. The model uses Kimi Delta Attention, Attention Residuals, and a sparse MoE design that activates 16 of 896 experts; Moonshot has not disclosed the active parameter count. Official API pricing is $0.30 per million cache-hit input tokens, $3 per million uncached input tokens, and $15 per million output tokens (Kimi technical blog, Kimi pricing docs). Moonshot's benchmark table puts K3 ahead of the compared models on BrowseComp, Program Bench, and OmniDocBench, but behind GPT-5.6 Sol on DeepSWE, Terminal-Bench 2.1, GPQA Diamond, and MMMU-Pro. An early independent run from Artificial Analysis scores K3 at 57 on its Intelligence Index, measures 62 output tokens per second, and flags 130M evaluation output tokens versus a 63M comparison median. That last number matters: K3's token price is attractive, but verbosity can erase part of the cost advantage.

Table of Contents

• Quick Verdict
• What Actually Shipped
• Architecture and Deployment Reality
• Kimi K3 API Pricing
• Kimi K3 Benchmarks
• Independent Reality Check
• Cost Per Workload
• API Contract and Migration
• Where Kimi K3 Wins and Loses
• Weights, Risks, and Falsifiers
• Final Recommendation
• FAQ
• About TokenMix
• Sources
• Related Articles

Quick Verdict

Kimi K3 is a credible frontier API today, but it is not yet a downloadable open-weight release.

The most important correction is the open-weight label. Moonshot calls K3 an open model and has committed to release the full weights by July 27. Until a real repository, license, checksums, and loadable files exist, the accurate status is weights announced, not yet published.

What Actually Shipped

The release is broader than a chat model: K3 ships across four products and one production API, while lower reasoning modes and the technical report remain pending.

This matters for procurement. API users can evaluate K3 now. Self-hosters cannot yet inspect the license or estimate real serving performance from released weights. The Kimi K3 integration guide should therefore be read as an API integration path until the promised repository exists.

Architecture and Deployment Reality

K3's 2.8T headline describes stored parameters, not per-token compute, because Moonshot disclosed expert routing but not active parameter count.

The raw storage floor is easy to calculate. At four bits per parameter, 2.8T x 4 / 8 equals about 1.4 TB decimal of weight data before quantization metadata, runtime buffers, expert routing, and the 1M-context KV state. That does not prove the final checkpoint size, but it explains why Moonshot recommends 64 or more accelerators.

KDA and Attention Residuals are more important than the parameter headline. Moonshot says KDA reduces long-sequence attention cost, while AttnRes retrieves earlier representations selectively instead of accumulating them uniformly. The architecture is plausible and technically specific. Real throughput, memory use, and hardware compatibility remain Likely, not Confirmed, until weights and serving code are public.

Kimi K3 API Pricing

K3 costs $3/$15 per million uncached input/output tokens, with a 90% cache-hit discount, but it is 3.1-3.7x more expensive than K2.6 on official RMB rates.

The dollar and RMB columns come from Moonshot's English launch page and Chinese API documentation. They are not currency conversions calculated by TokenMix. K3 does not use long-context price tiers: the same token rate applies across the 1M window.

The TokenMix row was read from the live model catalog on July 17 and should be checked again before production. Current availability is moonshot/kimi-k3; a catalog price is not a promise that future rates or cache accounting stay unchanged. For the older family, see the Kimi API pricing guide.

Kimi K3 Benchmarks

Moonshot's table shows a strong but uneven model: K3 wins three selected rows below and loses four to GPT-5.6 Sol or Fable 5.

Every number in this table is Vendor-reported, even where Moonshot cites an external leaderboard. The footnotes disclose different agent harnesses: KimiCode, Claude Code, and Codex are used depending on model and benchmark. That makes the table useful for direction, not a controlled one-variable comparison.

The honest verdict is narrower than the launch headlines. K3 looks especially competitive in browsing, document understanding, program synthesis, and long-horizon coding. It does not sweep scientific reasoning, terminal work, or general coding. Axios reached a similar boundary: the launch is important, but hours-old benchmark claims can overstate production reliability (Axios).

Independent Reality Check

Artificial Analysis independently places K3 near the frontier, but also measures two practical costs that Moonshot's benchmark chart does not foreground: speed and verbosity.

The 130M-token result is the cost warning. A model that generates roughly twice the median output can be cheap per token and still expensive per completed task. That does not mean K3 is wasteful on every workload; it means cost projections must use real usage logs rather than list prices alone.

Artificial Analysis currently labels K3 proprietary because no weights are public. Moonshot's July 27 commitment could change that classification. Until then, both statements can be true: Moonshot intends an open-weight release; the model is API-only today.

Cost Per Workload

At official rates, K3 costs $60 for 10M uncached input plus 2M output tokens and $384 for a 100M/20M workload with 80% cache hits.

Cost calculation 1: 10 x $3 + 2 x $15 = $60 for a small uncached K3 workload.

Cost calculation 2: with 80M cache hits and 20M fresh input, 80 x $0.30 + 20 x $3 + 20 x $15 = $384.

Cost calculation 3: a 1B-input, 200M-output uncached month is 1,000 x $3 + 200 x $15 = $6,000. If 90% of input hits cache, it falls to 900 x $0.30 + 100 x $3 + 200 x $15 = $3,570.

Those calculations still assume equal token consumption. If K3 emits twice as many tokens as another model on your workload, output cost can reverse the list-price advantage. Production evaluation should track cost per accepted task, not cost per million tokens.

API Contract and Migration

K3 is OpenAI-SDK compatible, but it is not a drop-in K2.6 session switch because reasoning history and sampling rules changed.

import os
from openai import OpenAI

client = OpenAI(
    api_key=os.environ["TOKENMIX_API_KEY"],
    base_url="https://api.tokenmix.ai/v1",
)

response = client.chat.completions.create(
    model="moonshot/kimi-k3",
    reasoning_effort="max",
    messages=[
        {"role": "user", "content": "Review this migration plan and list three failure modes."}
    ],
)

print(response.choices[0].message.content)

curl https://api.tokenmix.ai/v1/chat/completions \
  -H "Authorization: Bearer $TOKENMIX_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "moonshot/kimi-k3",
    "reasoning_effort": "max",
    "messages": [{"role": "user", "content": "Return a five-step canary plan."}]
  }'

Moonshot warns that K3 was trained with preserved thinking history. If a harness drops that history, or if a session switches from another model to K3 mid-conversation, quality can become unstable. Start a fresh K3 session and canary 100-300 real tasks before replacing the Kimi K2.6 route.

Where Kimi K3 Wins and Loses

K3 is strongest for long, tool-heavy work where 1M context and cache reuse matter; it is a poor default for short, latency-sensitive, tightly bounded tasks.

One official limitation deserves more attention than a benchmark delta: K3 can be excessively proactive on ambiguous tasks. For production agents, specify permission boundaries in the system prompt or AGENTS.md, require confirmation for destructive actions, and log every tool call.

Weights, Risks, and Falsifiers

The largest unresolved issue is not capability; it is whether the July 27 weight release arrives with a usable license, serving stack, and reproducible technical report.

The old version of this article predicted 3-4T parameters and a May release. Both are False after launch: K3 is 2.8T and released July 16. That correction is not cosmetic. It is why parameter rumors, prediction-market odds, and screenshots should never be elevated to Confirmed without an official model page.

Final Recommendation

Test Kimi K3 now if your workload is long-horizon coding, research, document analysis, or tool use. Do not migrate every K2.6 or K2.7 workload. K3 costs 3-4x more than K2.6 on official RMB token rates, only supports max reasoning today, and can generate unusually large outputs. Run a fresh-session canary, measure cost per accepted task, and revisit self-hosting only after the promised July 27 weights and license are public.

FAQ

Is Kimi K3 officially released?

Yes. Moonshot AI released Kimi K3 on July 16, 2026 across Kimi, Kimi Work, Kimi Code, and the Kimi API. The official API model ID is kimi-k3.

How many parameters does Kimi K3 have?

Kimi K3 has 2.8 trillion total parameters. Moonshot says 16 of 896 experts are activated, but it has not published the active parameter count. Earlier 3-4T claims were wrong.

What is Kimi K3 API pricing?

Official pricing is $0.30 per million cache-hit input tokens, $3 per million uncached input tokens, and $15 per million output tokens. Chinese documentation lists the same tiers as RMB 2, RMB 20, and RMB 100.

Is Kimi K3 open source or open weight?

Not as a downloadable model yet. Moonshot calls K3 open and promises full weights by July 27, 2026, but no official K3 checkpoint was visible on its Hugging Face organization when this article was verified. Check the files and license after release.

Does Kimi K3 beat GPT-5.6 Sol or Claude Fable 5?

Not across the board. Moonshot reports K3 wins on selected benchmarks such as BrowseComp and Program Bench, while GPT-5.6 Sol or Fable 5 leads DeepSWE, Terminal-Bench, FrontierSWE, GPQA, and other rows. Harness differences limit direct comparison.

Does Kimi K3 support a 1M-token context window?

Yes. The API documentation lists 1,048,576 tokens with no long-context price tier. Useful-context quality still requires independent testing on retrieval and long agent sessions.

Should I migrate from Kimi K2.6 to K3?

Only for workloads that benefit from stronger long-horizon reasoning, vision, or 1M context. K3 is roughly 3.08x more expensive for uncached input and 3.70x for output in official RMB pricing, so short or bulk tasks should remain on K2.6 or another cheaper route.

Can I call Kimi K3 through TokenMix?

Yes. The live TokenMix catalog lists moonshot/kimi-k3 with vision, tools, JSON, streaming, reasoning, structured output, and prompt caching support. Verify current availability and price in the catalog before production because routing and rates can change.

About TokenMix

TokenMix.ai provides access to Kimi, Claude, OpenAI, Gemini, DeepSeek, Qwen, GLM, and other models through one OpenAI-compatible endpoint at https://api.tokenmix.ai/v1. Check the live model catalog, pricing page, and API documentation before selecting a production route.

Sources

• Kimi K3: Open Frontier Intelligence - official architecture, benchmarks, availability, price, limitations
• Kimi K3 API Quickstart - API ID, reasoning contract, code, migration limits
• Kimi K3 API Pricing - RMB pricing and 1,048,576-token context
• Kimi K2.6 API Pricing - migration cost baseline
• Kimi API Model List - current platform availability
• Artificial Analysis: Kimi K3 - independent intelligence, speed, latency, verbosity
• Axios: China's open-weight Kimi model - independent launch coverage and caveat
• Xinhua: Kimi K3 launch - independent release-date confirmation
• OpenAI: GPT-5.6 - official comparison pricing
• Anthropic API Pricing - official Opus and Fable comparison pricing
• Moonshot AI Hugging Face Organization - public checkpoint status

Related Articles

• Kimi K2.6 Review: Pricing, Coding, and Benchmarks
• Kimi API Pricing 2026: K2.6, K2.5, and K2 Family
• Kimi K3 Developer Integration Guide
• Best Chinese AI Models 2026 Compared
• GLM-5.2 Review: 1M Context and Open Weights