Kimi K3: Moonshot's 2.8T Flagship, Open Weights, 1M Context

TL;DR

Published launch day, July 16, 2026; last updated August 21, 2026 with the July 27 weights release, the MXFP4 checkpoint details, and the first independent benchmark numbers.

Kimi K3 is Moonshot AI's flagship mixture-of-experts model, launched July 16, 2026. It has 2.8 trillion total parameters, a 1M-token context window, native visual understanding, and a new architecture: Kimi Delta Attention (hybrid linear attention) plus Attention Residuals. Moonshot reports 81.2 on FrontierSWE and 88.3 on Terminal-Bench 2.0, and positions K3 second only to Claude Fable 5 and GPT-5.6 Sol on GDPval-AA v2 (1687 Elo vs Opus 4.8's 1600). The weights shipped July 27 under Modified MIT as a native MXFP4 checkpoint.

What Is Kimi K3?

Kimi K3 is the successor to Moonshot AI's K2 line (K2.6 shipped April 2026 at 1T total / 32B active parameters; K2.7 Code followed in June and reached general availability inside GitHub Copilot). K3 nearly triples the total parameter count to 2.8T, extends context from 200K to 1M tokens, and adds native visual understanding for images and video. Thinking is always enabled, with reasoning traces exposed as separate deltas in the streaming API.

The launch was API-first: kimi-k3 on Moonshot's OpenAI-compatible platform, with the consumer Kimi app exposing two tiers, K3 Max and K3 Cluster Max. The weights followed eleven days later, on July 27. Moonshot is reportedly raising at a $31.5 billion valuation on the back of the release, up from $20 billion in May 2026, and is consolidating its platform around K3: the older kimi-k2.5 and moonshot-v1 model series are closed to new users and sunset entirely on August 31, 2026.

Before the official launch, a beta checkpoint codenamed "Kivine" ran anonymized on LM Arena, where early testers flagged two traits that carried into release coverage: elaborate, visually rich generations (interactive 3D scenes were a repeated example) and long runtimes on hard agent tasks.

Confirmed vs Leak Noise

K3 leaked for two days before it shipped, so early coverage mixed first-party facts with estimates. Most of the open questions have since resolved. The split as of August 21:

Architecture: Kimi Delta Attention at 1M Context

K3 is the first Moonshot flagship built on Kimi Delta Attention (KDA), the hybrid linear attention mechanism from the Kimi Linear paper (arXiv 2510.26692, October 2025). The design interleaves KDA linear-attention layers with periodic full-attention layers in a 3:1 ratio: three linear layers handle local sequence structure cheaply, one full-attention layer preserves global information flow. At matched scale in the paper, this cut KV-cache memory by up to 75% and delivered up to 6x decoding throughput at 1M-token context while matching or beating full-attention baselines on short-context, long-context, and RL-style post-training tasks. K3 pairs KDA with what Moonshot calls Attention Residuals.

The architecture is why the pricing is flat. Serving a 1M-token context with full attention means the KV cache, not the weights, dominates memory at long sequence length, which is why several competitors charge a long-context premium. Cut the cache 75% and the premium disappears from the cost structure. Whether serving economics at 2.8T actually work at $15/M output is Moonshot's bet.

At launch Moonshot published only the 2.8T total; K2.6's card had documented 1T total / 32B active, and leak coverage guessed K3 at 40-60B active, roughly double. Since July 27 the checkpoint itself is public, so the expert count, routing, and precision layout are inspectable in the Hugging Face config rather than argued from leaks. One serving-relevant detail from the repo: the checkpoint is natively MXFP4 quantization-aware trained (experts in MXFP4, activations in MXFP8), and a revision around August 5 moved four attention projections from BF16 to FP8 E4M3. Which revision and which precision path your provider serves is now a real fidelity question; Moonshot ships a public conformance harness, Kimi-Vendor-Verifier, for exactly this.

Benchmarks: Strong Numbers, All Self-Reported

Every K3 score below is Moonshot-reported. Launch-day skepticism on Hacker News centered on exactly this: open-weight-lineage models keep posting stunning vendor benchmarks, and the community suspects benchmark leakage into training data. Treat the table as the vendor's claim, not settled fact; independent harness results take a few weeks to land.

The BrowseComp claim deserves its own note: 91.2 in a single-agent setup with no context compression or context-management tricks. If that replicates, it means the 1M window plus KDA is doing the work that multi-agent orchestration and compaction pipelines normally do on long-horizon research tasks.

Independent signal, five weeks in

Independent testing settled about where day-one skeptics predicted: strong, one tier below the closed frontier. Artificial Analysis scores K3 60 on its Intelligence Index at max effort, tied with GLM-5.3 for the open-lineage lead and clearly above GLM-5.2 (53), and its AA-Briefcase agent evaluation puts K3 at 1543 Elo, second behind Claude Fable 5's 1574. That is below Moonshot's own GDPval-AA positioning but ahead of everything else with published weights. Developers consistently praise one thing the closed frontier does not offer: fully exposed reasoning traces, which one Hacker News commenter called "far, far more informative" than Fable's opaque summaries for debugging agent behavior.

The Token Economics: Always-On Thinking at $15/M

K3's pricing reads mid-tier until you account for how it generates. Thinking mode cannot be disabled, and reasoning_effort supports only max at launch (Moonshot says more levels are coming). So every request pays for a full reasoning trace at $15/M output tokens. This is the same failure mode that makes cheap-per-token models expensive per task: GLM-5.2 burns roughly 43K output tokens per Artificial Analysis Index task at max effort, and K3 launches with max as the only option.

Day-one Hacker News math framed it directly: $3/$15 with $0.30 cache hits matches Anthropic's Sonnet-series pricing, which is expensive for a model from the open-weight lineage, and commenters noted GLM-5.2 delivers similar coding quality at roughly one-third the per-token price. The counterpoint from K3's defenders: if the GDPval-AA positioning holds under independent testing, Sonnet-tier pricing for near-Fable capability is underpriced, not overpriced. Both can be true; which one matters depends on whether your workload needs the top tier.

The flat 1M-context pricing is a genuine differentiator. Anthropic charges a premium above 200K input tokens and Google tiers Gemini pricing by context length; Moonshot charges $3/M whether you send 4K or 900K tokens. For long-context-heavy workloads (repo-scale analysis, long agent traces, document piles) that flatness changes the calculus more than the headline rate does.

Kimi K3 API: Pricing and How to Call It

Morph serves Kimi K3 on GB300 NVL72 racks as morph-kimik3, on the same OpenAI-compatible API as the rest of the open-model lineup, at $2.80/M input and $14/M output, under Moonshot list pricing:

from openai import OpenAI

client = OpenAI(
    base_url="https://api.morphllm.com/v1",
    api_key="YOUR_MORPH_API_KEY",
)

resp = client.chat.completions.create(
    model="morph-kimik3",
    messages=[
        {"role": "user", "content": "Trace this race condition and propose a fix."},
    ],
)
print(resp.choices[0].message.content)

Moonshot's first-party API is OpenAI-compatible too: base URL https://api.moonshot.ai/v1, model kimi-k3. It supports streaming with separate reasoning and content deltas, structured JSON output with strict schema enforcement, tool calling with dynamic loading, vision inputs (base64 or uploaded file IDs), and a partial mode for prefix continuation. max_completion_tokens defaults to 131,072 and goes to 1,048,576.

Sources: Moonshot platform pricing (flat pay-as-you-go, no context-length tiering); OpenRouter model listing, which notes effective blended cost often lands 60-80% below list because of prompt caching. Since the weights opened, a dozen providers serve K3 with real spread in throughput and price; check the live provider tables before committing volume.

Where Are the Weights?

Out since July 27, 2026. Moonshot released Kimi K3 on Hugging Face (moonshotai/Kimi-K3) under a Modified MIT license, eleven days after the API launch, as a 96-shard checkpoint with an accompanying technical report. The K2-family precedent held. vLLM shipped day-zero support for the Kimi Delta Attention architecture, so the standard serving stacks run it from week one.

Two details matter before you deploy it. First, the checkpoint is natively MXFP4 quantization-aware trained: experts ship in MXFP4 with MXFP8 activations, so "the shipped checkpoint" is already a quantized artifact, and the community requants circulating on Hugging Face (NVFP4A16, W4AFP8, GGUF) are a second quantization on top. Second, the repo was revised around August 5, 2026, moving four attention projections from BF16 to FP8 E4M3; providers that pulled the launch revision and never re-synced are serving different weights than the current card. Moonshot publishes Kimi-Vendor-Verifier, a conformance harness that scores hosts on OCRBench, MMMU Pro Vision, BEAM-1M, and DeepSWE, to make exactly this visible.

The footprint is a different weight class than anything in the K2 line: even in native MXFP4 a 2.8T-total model is a multi-node deployment, and a hypothetical BF16 dequant would run roughly 5.6TB before KV cache. For comparison, GLM-5.2 at 753B is a 1.5TB BF16 deployment needing 8x H200-class GPUs for one node. Morph runs K3 on GB300 NVL72 racks; see the API section above.

Kimi K3 vs GLM-5.2 and DeepSeek V4

The relevant comparison set is the open-lineage frontier: GLM-5.2 (weights out, MIT), the API-only GLM-5.3 (same 743B base, weights gated at launch), DeepSeek V4 (weights out), and K3 (weights out since July 27). The honest summary five weeks in: K3 posts the strongest agentic numbers of the set and now has independent backing (Artificial Analysis Index 60, tied with GLM-5.3), at 2-10x the per-token price of the alternatives, with native vision as its unique capability.

Benchmark caveat: the Terminal-Bench versions differ (Moonshot reports 2.0, Z.ai reports 2.1), and every number in the table is self-reported by its vendor, so cross-model deltas of a few points are noise. What is not noise: the price gap. DeepSeek V4 Flash on Morph costs $0.09875/M input and $0.278/M output, roughly 2% of K3's output rate, and GLM-5.2 runs at about a third of K3's per-token price with weights you can hold. K3's case rests on the top-tier capability claims surviving independent testing.

For the previous Moonshot generation, see Kimi K2.5 and agent swarms. For the models Morph serves on custom codegen kernels, see Open Source Models: GLM-5.2, MiniMax M3, Qwen 3.5 397B, and DeepSeek V4 Flash.

Kimi K3: Pros and Cons

When to Use Kimi K3

Use K3 when the workload is long-horizon and context-heavy: single-agent research over large document sets (the BrowseComp setup), repo-scale analysis that genuinely needs several hundred thousand tokens in one window, or multimodal tasks where GLM-5.2 and DeepSeek V4 are disqualified for lacking vision. The flat 1M pricing makes it the cheapest way to actually use a million tokens of context in one call, even though its per-token rate is the highest of the open lineage.

Skip it, at least until independent benchmarks land, when the workload is high-volume codegen where per-task cost dominates. At $15/M output with always-on max reasoning, an agent loop that runs thousands of tasks a day costs an order of magnitude more on K3 than on DeepSeek V4 Flash or MiniMax M3, and the vendor-reported quality gap has not yet been independently priced. Teams running open models at real volume should also weigh serving quality: the same weights behave differently across hosts depending on quantization, speculators, and caching, which is where Morph's codegen-tuned serving earns its keep on the models it runs.

FAQ

What is Kimi K3?

Moonshot AI's flagship MoE model, launched July 16, 2026: 2.8T total parameters, 1M-token context, native vision, built on Kimi Delta Attention and Attention Residuals. Available as kimi-k3 via API, as K3 Max / K3 Cluster Max in the Kimi app, and as open weights on Hugging Face since July 27.

How much does the Kimi K3 API cost?

$3/M input tokens, $0.30/M on cache hits, $15/M output tokens on Moonshot, flat at any context length. Morph serves the same model as morph-kimik3 at $2.80/M input, $0.29/M cached, and $14/M output.

Is Kimi K3 open source?

Yes, since July 27, 2026: weights on Hugging Face (moonshotai/Kimi-K3) under a Modified MIT license, shipped as a native MXFP4 QAT checkpoint with a technical report and day-zero vLLM support.

Is Kimi K3 better than GLM-5.2?

On agentic benchmarks, yes: 88.3 vs 81.0 on Terminal-Bench (2.0 vs 2.1), and Artificial Analysis scores K3 60 against GLM-5.2's 53. On price, no: roughly 3x the per-token cost. The closer fight is GLM-5.3, which ties K3 at 60 on the AA Index at about one-fifth the price, but has no vision and gated weights.

Can I run Kimi K3 in Claude Code or Cline?

Any OpenAI-compatible client works against https://api.moonshot.ai/v1 with model kimi-k3, which covers Cline and similar tools directly. Keep the always-on reasoning in mind: agent loops that fire many small calls will pay max-effort reasoning on each one.

What happened to the 2.5T parameter figure?

That was the pre-launch leak number. The shipped spec in Moonshot's platform documentation says 2.8 trillion.