TL;DR: Kimi K3 — Moonshot AI’s 2.8-trillion-parameter model, released July 16, 2026 — is the first open-weight model to take #1 on a major frontend coding leaderboard, and its weights land on Hugging Face July 27. But the API costs $3/$15 per million tokens, it generates roughly twice the output tokens of comparable models, and it streams at 35 tokens/second. Sticker parity with Claude Sonnet 5 hides a real-world premium.

	Kimi K3 (API)	Claude Sonnet 5	DeepSeek V4 Pro
Best for	Frontend/UI generation, 1M-context jobs	Hardest multi-file refactors	Budget agent loops
Input / Output per 1M	$3.00 / $15.00 ($0.30 cache hit)	$2.00 / $10.00 intro → $3.00 / $15.00 Sep 1	$0.435 / $0.87
SWE-bench Verified	76.8% (vendor)	85.2%	80.6%
Frontend Code Arena	1,679 Elo — #1	1,631 (Fable 5; Sonnet unranked at top)	not in top tier
The catch	~2x verbosity, 35 tok/s, 51% hallucination rate	Intro pricing ends Aug 31	Trails badly on frontend

Honest take: Wire Kimi K3 into Cursor or Cline for frontend-heavy work and million-token context jobs, and keep Sonnet 5 or DeepSeek V4 Pro as your daily agent driver. K3 is a genuine milestone — but at Sonnet prices, double the token burn, and half the speed, the milestone is Moonshot’s, not your workflow’s.

What Moonshot actually shipped on July 16

Kimi K3 is the largest open-weight model ever announced: 2.8 trillion total parameters in a Mixture-of-Experts design with 896 experts, 16 of them active per token, built on Moonshot’s Kimi Delta Attention hybrid linear attention. Coverage puts active parameters per token in the 32–50B range — Moonshot hasn’t published one canonical figure. The context window is 1,048,576 tokens, input is natively multimodal (text, images, video), and reasoning is always-on at maximum effort — there is no “fast, shallow” mode. That last detail matters more than it sounds; it drives most of the cost math below.

The API went live July 16 at https://api.moonshot.ai/v1 with the model ID kimi-k3, OpenAI-compatible, with tool calls, JSON mode, structured output, and context caching. It’s also on OpenRouter as moonshotai/kimi-k3 at the same $3/$15 rate. The weights themselves are scheduled to publish on Hugging Face on July 27 — tomorrow, as this article goes up — under Moonshot’s Modified MIT license (standard MIT unless your product exceeds 100M monthly users or $20M/month revenue). As of July 24, independent trackers still classified K3 as hosted-only, so verify the repository actually exists before you plan anything around local inference.

One more launch-week wrinkle: demand was heavy enough that Moonshot paused new consumer app subscriptions on July 19. That pause applied to the Kimi app tiers (free up to $199/month), not the developer API — but it tells you the serving capacity is strained, which is consistent with the throughput numbers below.

The benchmarks: one real crown, several asterisks

The headline is legitimate: Kimi K3 debuted at #1 on the Frontend Code Arena with 1,679 Elo, ahead of Claude Fable 5 (1,631), GPT-5.6 Sol (1,618), and GLM-5.2 (1,587), ranking first in six of seven frontend domains (Fable 5 kept Gaming). This is a human-preference arena, not a vendor table, and it’s the first time an open-weight model has led a major coding leaderboard outright. For React/UI generation work, the crown is real.

The rest of the picture needs sorting into independent versus vendor-reported:

• Artificial Analysis Intelligence Index: 57 (independent) — AA ranks it #3, in the same band as Claude Opus 4.8 and GPT-5.5, behind only Fable 5 and GPT-5.6 Sol. For an open-weight model, unprecedented.
• Terminal-Bench 2.1: 88.3% (vendor-reported, run on Moonshot’s own KimiCode harness). That number would top the official tbench.ai chart — Fable 5 leads the native-harness board at 88.0% — but K3 isn’t on the official leaderboard yet, and as we covered in our Terminal-Bench analysis, harness choice alone can swing these scores by several points. Treat it as directional.
• SWE-bench Verified: 76.8% (vendor table). Here K3 trails the models it’s priced against: Sonnet 5 reports 85.2% and DeepSeek V4 Pro 80.6%. Backend bug-fixing is not where K3 wins.
• GDPval v2: 1,668 Elo (independent, AA) — a huge jump from K2.6’s 1,190, and past GLM-5.2, on long-horizon agentic knowledge work.

And then there’s the number Moonshot’s launch charts leave out. Artificial Analysis independently measured K3’s hallucination rate at 51% on fact-sensitive tasks — up from K2.6’s 39%. Overall knowledge reliability improved (AA-Omniscience index went from +6 to +18, accuracy from 33% to 46%), but the model got more confident about answering and more willing to fabricate when it shouldn’t. For coding, that pattern shows up as invented API surfaces and assured-sounding claims about library behavior. Whatever K3 tells you about a dependency, verify it — more so than you would with Sonnet.

There’s also a stranger data point worth one line: K3 has been caught identifying itself as Anthropic’s Claude in conversation, months after Anthropic publicly accused Moonshot, DeepSeek, and MiniMax of distillation campaigns. Moonshot hasn’t addressed it. Draw your own conclusions about training data provenance.

Wiring K3 into Cursor

Cursor’s native model picker includes Kimi K2.7 Code but not kimi-k3 as of July 26, so this is a BYOK job:

1. Create a key in the Moonshot console at platform.moonshot.ai (international billing; api.moonshot.cn is the China entrypoint).
2. In Cursor: Settings → Models → API Keys, paste the key into the OpenAI API Key field, enable Override OpenAI Base URL, and set it to https://api.moonshot.ai/v1.
3. Add kimi-k3 as a custom model name, then select it in Chat or Agent.

Sanity-check the key from your terminal before blaming Cursor for a misconfiguration:

curl https://api.moonshot.ai/v1/chat/completions \
  -H "Content-Type: application/json" \
  -H "Authorization: Bearer $MOONSHOT_API_KEY" \
  -d '{"model": "kimi-k3", "messages": [{"role": "user", "content": "Say ready."}]}'

A working key returns a JSON body with "model": "kimi-k3" and your reply under choices[0].message.content, plus a usage block you should get in the habit of reading — see the cost section. An invalid key returns a 401 with an error object.

The standing Cursor caveats apply, same as we documented for GLM 5.2: the base-URL override is global, so Cursor’s built-in models are unavailable while it’s active — flip it off when you want Sonnet back. Tab autocomplete always runs on Cursor’s own models regardless of BYOK. And if you’d rather not open another billing account, OpenRouter serves moonshotai/kimi-k3 through its standard OpenAI-compatible endpoint at identical pricing.

A trap we hit in the docs: the 1M context window is gated by your Moonshot API tier. Early setup guides report lower tiers cap K3 at 256K, with the full 1M reserved for higher-spend tiers. If your million-token dream was the reason you signed up, check your tier’s limit in the console before you pipe a whole monorepo into a prompt and get a confusing overflow error at 256K.

Wiring K3 into Cline

Cline is easier — it ships a first-party Moonshot provider, no OpenAI-compatibility shim needed:

1. Open Cline’s settings and select Moonshot as the API provider.
2. Choose the entrypoint (api.moonshot.ai international, api.moonshot.cn China).
3. Paste your key and set the model to kimi-k3.

Cline’s plan/act split pairs well with a cheaper model for the act phase. Running K3 as the planner and DeepSeek V4 Pro or Gemini 3.6 Flash as the executor keeps K3’s frontend judgment where it pays and takes the 17x-cheaper output rate everywhere else.

Claude Code users aren’t left out: Moonshot ships an Anthropic-compatible endpoint, so ANTHROPIC_BASE_URL=https://api.moonshot.ai/anthropic plus your Moonshot key routes Claude Code’s traffic to K3 — the same pattern Moonshot documented for Kimi K2.7 Code.

The cost math nobody puts in the launch post

On paper, K3 at $3/$15 costs exactly what Sonnet 5 will cost from September 1 (Sonnet is $2/$10 intro through August 31). In practice, three measured facts break the parity:

K3 talks twice as much. Across Artificial Analysis’s Intelligence Index evaluation, K3 generated 130M output tokens where the median model generates 63M. On agentic tasks AA measured ~120K output tokens and 83 turns per task, versus 42K tokens and 54 turns for K2.6. Always-on maximum reasoning means you pay for every thinking token at $15/M.

K3 is slow. AA measured 35 tokens/second against a 78 tok/s median. At that rate, a 120K-output-token agentic task spends nearly an hour just generating text. For an overnight batch job, fine. For interactive Cursor work, you will feel it on every request.

The verbosity compounds the price. Rough per-task math, using AA’s measured output profiles:

Backend	Output tokens/task (measured)	Output cost/task	Sticker price/M out
Kimi K3	~120K	~$1.80	$15.00
Kimi K2.6/K2.7-class	~42K	~$0.17 (at K2.7’s $4/M)	$4.00
DeepSeek V4 Pro	mid-pack	~$0.05–0.10	$0.87

Context caching softens the input side — 83 turns per task means enormous repeated context, and $0.30/M cache-hit pricing (a 90% discount) is genuinely useful there. But output tokens can’t be cached, and output is where K3’s meter runs hot. The Decoder called K3 the end of the ultra-cheap Chinese AI era, and the numbers back that up: K3 costs 3x–3.75x its own sibling K2.7 Code and roughly 17x DeepSeek V4 Pro on output. Our cost comparison table has the full market view.

What July 27 actually changes

The weights drop is the story everyone will run tomorrow, so here’s the sober version. A 2.8T-parameter model is about 1.4TB of weights even at aggressive 4-bit quantization (community discussion centers on MXFP4). That is not “grab a GGUF” territory — it’s a multi-node H200 rack, or a cluster of three-plus 512GB Mac Studio M3 Ultra machines wired together. If you want to experiment without owning the hardware, renting GPU capacity by the hour on RunPod is the realistic path; for what actually fits on hardware humans own, see runaihome.com’s local model VRAM guide, and watch the FOSS side at aifoss.dev for self-hosting coverage as the weights land.

What the drop does change: hosting competition. Once weights are public, Together, Fireworks, DeepInfra and friends can serve K3 and undercut Moonshot’s first-party pricing — exactly what happened with GLM-5.2, which aggregators now serve at less than half of Z.ai’s $1.40/$4.40 list. If K3’s third-party prices land meaningfully below $3/$15, the cost verdict above gets rewritten. It also means independent benchmark runs on neutral harnesses, which will settle the Terminal-Bench question Moonshot’s self-reported 88.3% can’t.

Verdict

Kimi K3 is the most capable open-weight model ever released, and for frontend generation it isn’t just competitive — it leads, against every closed frontier model, on a human-preference arena. If your work is UI-heavy, wire it into Cline behind the native Moonshot provider and use it where it’s strongest. And the 1M context window at $0.30/M cached input makes it a legitimate tool for whole-repo comprehension jobs that choke smaller-context models.

As a default backend, it loses to both of its price neighbors. Sonnet 5 beats it by 8+ points on SWE-bench Verified at the same September price, twice the speed, and roughly half the output tokens per task. DeepSeek V4 Pro delivers most of the agentic capability at a twentieth of the effective cost. Between the verbosity tax, the 35 tok/s wait, and a 51% measured hallucination rate that Moonshot’s marketing doesn’t mention, K3 is a specialist and a preview of where open models are going — not the thing to point your Cursor agent at on Monday morning. Re-evaluate after July 27 when third-party hosts publish their prices.

FAQ

What’s the exact model ID and endpoint? kimi-k3 at https://api.moonshot.ai/v1 (OpenAI-compatible; api.moonshot.cn for China). On OpenRouter it’s moonshotai/kimi-k3. Moonshot also exposes an Anthropic-compatible endpoint for Claude Code at /anthropic.

Can I run Kimi K3 locally after July 27? Only with data-center hardware. 2.8T parameters is ~1.4TB of weights at 4-bit — multi-GPU-node or large Mac-cluster territory. For local coding models that fit in real VRAM budgets, a 24–48GB card running Qwen3-Coder or Devstral remains the practical answer.

Is Kimi K3 actually #1 at coding? It’s #1 on the Frontend Code Arena (human-preference, independent) — a real first for open weights. On SWE-bench Verified (76.8%, vendor-reported) it trails Sonnet 5 and DeepSeek V4 Pro, and its 88.3% Terminal-Bench claim comes from Moonshot’s own harness and isn’t on the official leaderboard yet.

Why is my K3 context capped at 256K when the model supports 1M? The 1M window is gated by Moonshot API tier. Lower-spend tiers reportedly get 256K; check your tier limits in the platform console.

Does the July 19 subscription pause affect the API? No — Moonshot paused new consumer app subscriptions after launch demand spiked. API key creation and billing were unaffected, though the capacity strain likely contributes to the modest 35 tok/s serving speed.

Is K3 cheaper than Claude Sonnet 5? Not really. The sticker matches Sonnet’s September rate ($3/$15) and costs more than Sonnet’s intro rate ($2/$10 through August 31) — and because K3 emits roughly twice the output tokens per task, its effective per-task cost runs meaningfully higher than either Sonnet number. Budget 2–3x the sticker math for agent loops.

Sources

• China’s Moonshot AI releases Kimi K3, the largest open-source model ever — VentureBeat
• Kimi K3 beats Claude Fable 5 in Frontend Code Arena benchmark — Tom’s Hardware
• Kimi K3 achieves #3 in the Artificial Analysis Intelligence Index — Artificial Analysis
• Kimi K3 — Intelligence, Performance & Price Analysis, Artificial Analysis
• Kimi K3 Open Weights Drop July 27: Near-Frontier Coding, Undisclosed Hallucination Risk — TechTimes
• Kimi’s open model K3 nears GPT-5.6 Sol and Fable 5 while signaling the end of super cheap Chinese AI — The Decoder
• Kimi K3 — API Pricing & Benchmarks, OpenRouter
• Kimi K3 quickstart — Kimi API Platform docs
• Moonshot provider configuration — Cline docs
• Frontier model API pricing, July 2026 — Developers Digest
• Kimi K3 identifies itself as Anthropic’s Claude — Wccftech

Last verified July 26, 2026. Pricing, tier limits, and the weights-release status change fast — check platform.moonshot.ai before committing a team budget.