# Kimi K3: Moonshot AI's Frontier Model, Assessed

Kimi K3 is a 2.8-trillion-parameter mixture-of-experts model, released by Moonshot AI on 16 July 2026. On the independent Artificial Analysis Intelligence Index it placed fourth of the field at 57, above Claude Opus 4.8 and GPT-5.5, at an API rate several times below either. That placement is what made the release consequential.

Two questions sit unresolved behind the score. The weights are not public despite the open-weight framing, and the company carries an unresolved allegation from Anthropic about how its models were trained. The capability is measurable today. The openness and the provenance are not.

## The model

Kimi K3 is a 2.8-trillion-parameter mixture-of-experts model with a 1,048,576-token context window and always-on reasoning. Moonshot describes it as "the world's first open 3T-class model." That description carries an important qualification. As of mid-July the weights are not public. Moonshot has committed to releasing them by 27 July 2026, and its K2 lineage supports the promise, but Artificial Analysis currently classifies K3 as a proprietary model and answers its own question, "Is Kimi K3 open source?", with "No, weights are not publicly available." The accurate reading is that K3 is an announced open-weight model with a release date, available today only through the API.

| Attribute | Value |
| --- | --- |
| Parameters | 2.8T total, 16 of 896 experts active per token (Stable LatentMoE) |
| Context window | 1,048,576 tokens, flat pricing across the window |
| Modalities | Text, image, and video input (per Moonshot docs; AA lists text and image only) |
| Reasoning | Always on; `reasoning_effort` currently max only |
| Weights | Modified MIT expected, due ~27 July 2026 (not yet released) |
| API price (per 1M) | $3.00 input cache-miss, $0.30 input cache-hit, $15.00 output |

The architecture is the substantive novelty. Moonshot builds K3 on two components it names directly, Kimi Delta Attention, a hybrid linear attention mechanism that governs information flow across the sequence, and Attention Residuals, which retrieve representations selectively across model depth rather than accumulating them uniformly. The company claims roughly 2.5 times the scaling efficiency of Kimi K2 from this design combined with its training recipe. Serving runs on Mooncake, Moonshot's disaggregated inference platform, which processes 100 billion tokens per day and is credited with the cache-hit rate above 90 percent in coding workloads that makes the $0.30 cached-input price feasible. A full technical report is promised with the weights; until then these claims rest on blog prose rather than a paper. The caching mechanics and the tool-calling design are documented on the [technical page](https://comparedge.com/tools/kimi/api).

## Independent benchmarks

Artificial Analysis places Kimi K3 fourth on its Intelligence Index, at 57, against a field average near 30. The model sits above Claude Opus 4.8 and GPT-5.5, and below the two closed leaders it was measured against, Claude Fable 5 and GPT-5.6 Sol.

| Model | AA Intelligence Index |
| --- | --- |
| Claude Fable 5 | 60 |
| GPT-5.6 Sol | 59 |
| **Kimi K3** | **57** |
| Claude Opus 4.8 | 56 |
| GPT-5.6 Terra | 55 |
| GPT-5.5 (xhigh) | 55 |
| Grok 4.5 | 54 |
| Claude Sonnet 5 | 53 |
| GLM-5.2 | 51 |
| Gemini 3.5 Flash | 50 |
| DeepSeek V4 Pro | 44 |

The composite score understates where the model is strongest. Its AA Coding Index is 76.2 and its Agentic Index is 50.1, both among the top few families measured. It ranks first of the field on long-context reasoning, which is consistent with the million-token window. On raw knowledge it is mid-pack, scoring 46 percent on the omniscience evaluation against Fable 5's 61. Two operational figures qualify the picture. Output speed is 62 tokens per second, below the field median of 71, and the model is unusually verbose, generating 130 million output tokens to complete the full index against a 63-million-token field average. Verbosity is a benchmark curiosity and a cost problem at once, and it recurs below.

## The vendor's own sheet

Moonshot's published benchmark table is more favorable than the independent index, as vendor sheets are, and the company is candid about the ceiling. It states that K3 "trails the most powerful proprietary models, Claude Fable 5 and GPT 5.6 Sol," while leading the remainder of its test set. The coding and agentic rows are the ones that matter for the model's stated use cases.  

![Editorial benchmark comparison visualization showing Kimi K3 performance across coding and agentic AI benchmarks versus Claude Fable 5, GPT-5.6 Sol, Opus 4.8, GPT-5.5, and GLM-5.2. The infographic highlights vendor-reported benchmark scores, methodology notes about context compaction, and Kimi K3's position as a top-tier frontier model behind the leading proprietary AI systems.](https://cdn.hashnode.com/uploads/covers/69e67cdfc9501dd01024f1a7/aadeda1e-7809-4772-9d23-ab394c968a3d.png align="center")

| Coding benchmark | Kimi K3 | Fable 5 | GPT-5.6 Sol | Opus 4.8 | GPT-5.5 | GLM-5.2 |
| --- | --- | --- | --- | --- | --- | --- |
| Terminal Bench 2.1 | 88.3 | 84.6 | 88.8 | 84.6 | 83.4 | 82.7 |
| FrontierSWE | 81.2 | 86.6 | 71.3 | 66.7 | 64.9 | 67.3 |
| SWE Marathon | 42.0 | 35.0 | 39.0 | 40.0 | 14.0 | 13.0 |
| Program Bench | 77.8 | 76.8 | 77.6 | 71.9 | 70.8 | 63.7 |

| Agentic benchmark | Kimi K3 | Fable 5 | GPT-5.6 Sol | Opus 4.8 | GPT-5.5 |
| --- | --- | --- | --- | --- | --- |
| BrowseComp | 91.2 | 88.0 | 90.4 | 84.3 | 84.4 |
| DeepSearchQA (F1) | 95.0 | 94.2 | n/a | 93.1 | n/a |
| Automation Bench | 30.8 | 29.1 | 29.7 | 27.2 | 22.7 |
| MCP Atlas | 84.2 | 84.7 | 83.6 | 83.6 | 82.8 |

One methodological detail deserves attention. The headline BrowseComp figure of 91.2 uses a context-compaction strategy triggered at 300,000 tokens, not the full context window. Run with the complete million-token window and no compaction, Moonshot reports 90.4 for the same evaluation. When the independent index and the vendor sheet agree on the shape, strong, top-five, behind the two best US models, that shape is trustworthy. The vendor's individual numbers are the optimistic reading.

Moonshot also discloses limitations that a benchmark chart does not surface. K3 was trained in a preserved-thinking-history mode, and the company warns that if an agent harness fails to return the full reasoning history, or a session is switched to K3 mid-stream from another model, "generation quality may become highly unstable." It further notes a tendency toward excessive proactiveness on ambiguous instructions and, plainly, "a noticeable gap in user experience compared with Claude Fable 5 and GPT 5.6 Sol."

## Preference leaderboards

The claim that circulated most widely at launch, that K3 beat both closed leaders, is true on one board and false on another. This distinction is the difference between an accurate result and an overstated one.

| Arena board | Kimi K3 | Claude Fable 5 | GPT-5.6 Sol |
| --- | --- | --- | --- |
| Frontend Code Arena | #1 (1,679 Elo) | #2 (1,631) | #3 (1,618) |
| Main Text Arena | #6 (1,500, preliminary) | above K3 | above K3 |

On the Frontend Code Arena, K3 debuted first, a 17-place jump over Moonshot's prior model, with a 76 percent pairwise win rate on frontend tasks. On the general Text Arena it placed sixth, behind both leaders. The viral framing that omitted the board qualifier turned a genuine frontend-code result into an inaccurate general one. Arena's chief executive, Anastasios Angelopoulos, called it "the single biggest release of the year" and said it "marks the moment that OSS Chinese models have surpassed US models," a named opinion about a preference-vote board rather than a settled measure of capability.

## Economics

The published rate is $3.00 per million input tokens, $0.30 on a cache hit, and $15.00 per million output. Blended, that is roughly $2.31 per million, several times below the two models ranked above it. Prompt caching is automatic, with no cache identifiers or expiry to manage, and the effect on input cost is large.

| Model | Index | Input / 1M | Output / 1M |
| --- | --- | --- | --- |
| Claude Fable 5 | 60 | $10.00 | $50.00 |
| GPT-5.6 Sol | 59 | $5.00 | $30.00 |
| **Kimi K3** | **57** | **$3.00** ($0.30 cached) | **$15.00** |
| Claude Opus 4.8 | 56 | $5.00 | $25.00 |
| GPT-5.6 Terra | 55 | $2.50 | $15.00 |
| DeepSeek V4 Pro | 44 | $0.435 | $0.87 |

![Kimi K3 AI pricing infographic comparing input and output token costs with Claude Fable 5, GPT-5.6 Sol, Claude Opus 4.8, GPT-5.6 Terra, and DeepSeek V4 Pro. The diagram highlights automatic prompt caching, effective input cost reduction, output token pricing, and cache-hit rate impact on total inference costs.](https://cdn.hashnode.com/uploads/covers/69e67cdfc9501dd01024f1a7/ca547736-7ccd-42cc-adc7-e1592487184e.png align="center")

  
Input becomes cheap when context repeats. A 50,000-token system prompt costs $0.15 the first time and $0.015 on a cache hit. The cost that does not compress is output. Because reasoning is always on and the model is verbose, the $15.00 output rate dominates most real bills. Gavin Baker of Atreides Management estimated that K3 runs 50 to 70 percent more expensive in practice than its headline pricing implies, once reasoning overhead is counted. There is no Batch discount for K3 at launch, and the pricing marks a shift in the category more broadly. Kimi K2.6 cost $0.95 per million input; at $3.00, K3 is priced into Western territory, which The Decoder characterized as the end of super-cheap Chinese AI.

| Cache-hit rate | Effective input / 1M |
| --- | --- |
| 0% (novel prompts) | $3.00 |
| 90% (vendor's coding claim) | $0.57 |
| 93% (observed launch week) | $0.489 |

The verbosity multiplier, the missing Batch discount, and the break-even against an Opus-tier provider are worked through on the [cost guide](https://comparedge.com/tools/kimi/cost-guide); the five consumer membership tiers, from a free plan to $199 per month, are itemized on the [pricing page](https://comparedge.com/tools/kimi/pricing). The live rate card is also exposed through ComparEdge's public [API and MCP endpoint](https://comparedge.com/api-docs) for programmatic cost modeling.

## The company behind it

Moonshot AI was founded in March 2023 by three Tsinghua University graduates and named 月之暗面, the dark side of the moon, after the Pink Floyd album. The founding team is stronger than the company's age suggests. Yang Zhilin, the chief executive, completed a Carnegie Mellon doctorate in four years and co-authored XLNet and Transformer-XL. Zhou Xinyu, the chief technology officer, co-authored ShuffleNet. Wu Yuxin, who leads research, built Detectron2 at Meta. Moonshot is counted among China's "six AI tigers" and employs roughly 300 people.

The funding record explains why the roadmap should be taken seriously.

| Date | Amount | Valuation | Lead |
| --- | --- | --- | --- |
| 2023 (seed) | ~$60M | $300M | n/a |
| Feb 2024 | $1B | $2.5B | Alibaba (~36% stake) |
| Aug 2024 | $300M | $3.3B | Tencent, Gaorong |
| ~Oct 2025 | ~$600M | $3.8B | IDG Capital |
| Early 2026 | $700M | $10B | n/a |
| **May 2026** | **$2B** | **$20B** | **Meituan Long-Z** |
| Jun 2026 | ~$2B sought | ~$30B target | (talks) |

TechCrunch counted $3.9 billion raised over six months, and annualized revenue passed $200 million in April 2026, most of it from the API. Alibaba and Tencent are repeat backers. The company is unwinding an offshore structure to prepare a Hong Kong listing. A firm this funded, growing this quickly, is a commercial operation rather than a research demonstration, and the $20 billion valuation is the closed figure; the $30 billion is a forward signal from reporting on talks.

## Governance, training data, and the distillation question

This is where the model's strong numbers meet its weakest ground. Three issues bear on any adoption decision.

The first is certification. Moonshot publishes none. There is no SOC 2, ISO 27001, or equivalent in the terms, privacy policy, or documentation. For a regulated buyer that is decisive regardless of benchmark performance. The second is data handling. On the standard API, content may be used to improve the models unless the customer signs an enterprise agreement, so the default runs toward training on submitted data rather than away from it. Data residency is the one area with a clean answer: the global platform stores and processes data in Singapore under Singapore law, explicitly not mainland China, while the China platform operates separately under the Beijing entity.

The third is the training-provenance allegation, which should be reported precisely because it is serious and unresolved. In February 2026 Anthropic accused three Chinese firms, DeepSeek, Moonshot, and MiniMax, of using fraudulent accounts to extract Claude's reasoning and tool-use traces for training. Anthropic's figures across the three companies were roughly 24,000 fraudulent accounts and more than 16 million exchanges, with the Moonshot-specific share reported at hundreds of accounts and 3.4 million exchanges. The primary source is Cade Metz in The New York Times, 23 February 2026. There is no public ruling or settlement, and Moonshot has not addressed the allegation on record. A separate claim that surfaced after the K3 launch, that the model reproduces Claude output nearly word for word, is weaker and was largely dismissed within the technical community itself, on the grounds that superficially similar output across unrelated model families is common. The February allegation stands as a credible, dated, unproven claim; the July "proof" does not carry the same weight.

## Reception

The commentary split along predictable lines. Holger Mueller of Constellation Research called K3 "the largest open-weights model we've ever seen" and "another DeepSeek moment." Patrick Moorhead judged the response "an overreaction shockingly similar" to DeepSeek's. David Sacks, who chairs the White House science and technology council, called the frontend-code result "concerning," and Bill Ackman replied "Agreed." Prediction markets stayed unmoved, with Polymarket giving Anthropic better than a two-thirds chance of holding the year-end top position. The most durable argument came from the developer community rather than the analysts: that the only lasting US advantage is the hardware gap created by export controls, and that if model quality remains uncoupled from hardware, that advantage erodes, while if it recouples in 2027, a gap reopens. Several near-term Chinese releases, GLM 5.5, MiniMax Pro, and Qwen 4, suggest K3's lead, where it exists, may be brief.

## Assessment

Kimi K3 is the strongest price-for-intelligence position at the frontier, and the qualifications attach to timing and trust rather than to capability. For a cost-sensitive team doing high-volume coding or agentic work, on data it controls, the model is among the best available at any price, and materially cheaper than the two systems ranked above it. Teams whose interest depends specifically on the open weights should wait for 27 July, since today the offering is API access rather than a model to self-host. Teams that require SOC 2 or ISO 27001, a no-training-by-default posture, or the strongest hard-science reasoning are better served by an Opus-tier provider, which costs more per token and answers the questions an auditor will ask.

The capability is real and independently confirmed. The openness is a date on the calendar. The provenance is contested. Each should be weighed on its own terms. The complete assessment, with every figure linked to its source, is on the [ComparEdge Kimi K3 review](https://comparedge.com/tools/kimi).