# Kimi K3 Explained: China's New AI Heavyweight.

- Canonical page: https://mattfarmer.ai/kimi-k3
- Author: Matt Farmer
- Published: 2026-07-17
- Last verified: 2026-07-17
- Evidence cutoff: 2026-07-17
- Status: Kimi products and API are live. Full weights and a technical report are promised by July 27, 2026, but the checkpoint and license were not verified at this cutoff.

## Plain-English Summary

Moonshot AI released a Chinese AI model that scores close to GPT-5.6 Sol and Claude Fable 5, costs less per token, and is promised to become open weight on July 27, 2026.

The important word is **close**. The July 17 Artificial Analysis snapshot scored K3 at 57, GPT-5.6 Sol at 59, and Claude Fable 5 at 60. K3 did not win that broad independent composite, but it landed close enough to belong in the same serious model conversation.

K3 is usable now through Kimi products and the official API. The open-weight checkpoint, license, and technical report were still promises—not downloads—at the evidence cutoff.

## Why K3 Matters

K3 combines native image understanding, a one-million-token context window, long-horizon agent work, consumer products, coding tools, and a paid API. Moonshot describes a 2.8-trillion-parameter sparse model that activates 16 of 896 experts for each token.

The promised open-weight release could give researchers and companies more control than a closed API alone. It would not make the full model easy to run: Moonshot recommends at least 64 accelerators for full deployment.

## How Close Is K3 to GPT-5.6 and Claude?

### Independent Intelligence Index

Artificial Analysis Intelligence Index v4.1, captured July 17, 2026. Higher is better. This is an independent composite, not a percentage score.

| Model | Score |
| --- | ---: |
| Claude Fable 5 | 60 |
| GPT-5.6 Sol | 59 |
| Kimi K3 | 57 |

K3 is three points behind Fable and two behind Sol in this snapshot.

### Preliminary Arena WebDev Snapshot

Arena WebDev Overall, captured July 16, 2026. K3's **1679 +/-17 result was Preliminary** and can change as more votes arrive.

| Model | Preliminary WebDev snapshot |
| --- | ---: |
| Kimi K3 | 1679 ±17 (Preliminary) |
| Claude Fable 5 | 1631 |
| GPT-5.6 Sol | 1618 |

K3 led this specific web-development snapshot. That supports a visual-coding use case; it does not prove K3 wins every workload.

## API Pricing

Published price per million tokens:

| Model | Cache-miss input | Output | Cached input |
| --- | ---: | ---: | ---: |
| Kimi K3 | $3 | $15 | $0.3 |
| GPT-5.6 Sol | $5 | $30 | Provider-specific |
| Claude Fable 5 | $10 | $50 | Provider-specific |

K3 has the lowest published input and output rates in this three-model comparison. Provider caching rules, output volume, turns, retries, tools, taxes, and cleanup determine the completed-job cost.

Three examples from the supplied worksheet:

- **Quick chat: about $0.02.** One turn; no tools or retries.
- **Coding with a familiar project: about $0.36.** 90% of 100K input cached; 20K output/reasoning.
- **Long research task: about $1.69.** 75% of 500K input cached; tools and web fees excluded.

Formula:

`((cached input / 1M x $0.30) + (uncached input / 1M x $3.00) + (output / 1M x $15.00)) x turns x retry multiplier + tool fees`

## What Ordinary People Can Do With K3

- **Build websites and software.** K3 is designed for long coding sessions, visual website work, debugging, and multi-step agent tasks.
- **Research very long documents.** Its one-million-token context can hold large collections of reports, transcripts, reference material, and code.
- **Create office work.** Moonshot demonstrates documents, presentations, spreadsheets, dashboards, and other editable business artifacts.
- **Understand images.** Native visual input supports screenshot, chart, document, design, and visual-coding tasks.

These are reasons to test K3, not guarantees that every output will be correct.

## What Open Weight Means

Open weights means the trained model files become available under whatever terms the license permits. Open source is a broader claim involving code, documentation, rights, and the exact license.

Verified timeline at the evidence cutoff:

1. **July 16:** K3, consumer products, and API launched.
2. **July 17:** Evidence and Moonshot's public repositories were checked.
3. **July 27:** Full weights and a technical report were promised.

At cutoff, the checkpoint, license, technical report, third-party hosts, and independent self-host results were not verified. Do not call K3 open source or downloadable until those artifacts can be inspected.

## The Catch

1. **Close is not first.** K3 remained behind Fable and Sol in the independent composite.
2. **Cheap tokens can become an expensive job.** Artificial Analysis measured 130 million output tokens across its index run.
3. **Hosted API privacy needs review.** Zero retention and no-training-by-default were not verified as standard promises.
4. **The open-weight story is still pending.** The checkpoint and license were not available at the July 17 cutoff.

## Technical Evidence

### Release Facts

| Fact | Value | Evidence |
| --- | --- | --- |
| Launch | July 16, 2026 | Official |
| API model ID | kimi-k3 | Official |
| Parameters | 2.8T total (vendor-stated) | Official |
| Expert routing | 16 active of 896 | Official |
| Context | 1M tokens | Official |
| Input | Text, images, and documented video-file input | Official |
| API price | $0.30 cached / $3 miss / $15 output per 1M | Official |
| Reasoning effort | API: max only; Kimi Code docs: low / high / max | Official |
| Weights | Promised by July 27; not shipped at cutoff | Pending |

### Independent Measurements

- **57 AA Intelligence Index.** Score 57, ranked fourth of 187 models at the July 17 capture.
- **62 Output tokens / second.** Artificial Analysis measured 62 output tokens per second.
- **1.99 Seconds to first token.** Artificial Analysis measured a 1.99-second time to first token.
- **130000000 Output tokens in AA run.** The Intelligence Index run generated 130 million output tokens, making output volume part of the cost story.

### Full Twelve-Model Comparison

| Model | AA intelligence snapshot | Context | Published API pricing per 1M tokens | Weights | Strongest fit | Main caveat |
| --- | ---: | ---: | --- | --- | --- | --- |
| Kimi K3 | 57 | 1M | $0.30 cached / $3 input / $15 output | Promised July 27; not shipped | Long-horizon agents; WebDev; visual coding; research workflows | Verbose; 62 t/s; preserved-thinking contract; 64+ accelerator guidance |
| Claude Fable 5 | 60 | 1M | $10 input / $50 output | Closed | Hardest reasoning and agent work | Highest cost; fallback/access/retention considerations |
| GPT-5.6 Sol | 59 | 1.05M | $5 input / $30 output | Closed | Premium coding; agents; vision; lower token use | Higher sticker price and product-specific long-context rules |
| Claude Opus 4.8 | Not reported | 1M | Official provider pricing | Closed | Mature Anthropic agent baseline | Behind current Fable generation |
| GPT-5.5 | Not reported | 1.05M | Official provider pricing | Closed | Mature Codex-harness baseline | Behind current Sol generation |
| GLM-5.2 | 51 | source-specific | Provider dependent | Available | Current highest AA-scored downloadable open weights | Lower AA score than K3; provider variance |
| Kimi K2.7 Code | Not reported | 256K | $0.19 cached / $0.95 input / $4 output | Available | Mature and lower-cost Kimi coding | Lower peak capability and context |
| Kimi K2.6 | Not reported | 256K | $0.55 input / $2.65 output | Available | Fast general Kimi mode | Prior generation |
| Gemini 3.1 Pro | Not reported | 1M | $2 input / $12 output | Closed | Broad multimodal reasoning | Preview/surface status and not in Moonshot main table |
| Grok 4.5 | Not reported | 500K | $2 input / $6 output | Closed | Coding-agent value and live X research | Smaller context and regional access |
| MiniMax M3 | 44 | source-specific | Provider dependent | Available | Open-weight agent/value option | Below K3 and GLM-5.2 AA index |
| DeepSeek V4 Pro | 44 | source-specific | Provider dependent | Available | Open-weight reasoning/value option | Max-effort and provider variance |

### Full 37-Row Benchmark Matrix

Moonshot's launch table mixes harnesses, effort levels, vendor tests, internal tests, and third-party results. These rows keep the original configuration and comparability caveat instead of blending unlike evidence into one score.

| Category | Benchmark | Metric | K3 | Fable 5 | GPT-5.6 Sol | Opus 4.8 | GPT-5.5 | GLM-5.2 | K3 result | Scope | Harness | Evidence | Comparability caveat |
| --- | --- | --- | ---: | ---: | ---: | ---: | ---: | ---: | --- | --- | --- | --- | --- |
| Coding | DeepSWE | score | 67.5 | 70 | 73 | 59 | 67 | 46.2 | Trails | Vendor test | KimiCode | Vendor Test | Official leaderboard figures use multiple harnesses; K3 is 67.3 under mini-SWE-agent. |
| Coding | Program Bench | pass rate | 77.8 | 76.8 | 77.6 | 71.9 | 70.8 | 63.7 | Leads | Vendor test | KimiCode | Vendor Test | GLM score comes from vendor blog; others from Vals AI per Moonshot. |
| Coding | Terminal Bench 2.1 | score | 88.3 | 84.6 | 88.8 | 84.6 | 83.4 | 82.7 | Trails | Vendor test | KimiCode | Vendor Test | Other models use best reported harness; not a pure model comparison. |
| Coding | FrontierSWE | dominance | 81.2 | 86.6 | 71.3 | 66.7 | 64.9 | 67.3 | Trails | Vendor test | KimiCode | Vendor Test | Harness differs; dominance recomputed from raw scores on July 16. |
| Coding | SWE Marathon | score | 42 | 35 | 39 | 40 | 14 | 13 | Leads | Vendor test | Claude Code | Vendor Test | K3/Fable/Opus use Claude Code; Sol uses Codex; GLM from vendor. |
| Coding | PostTrain Bench | score | 36.6 | 41.4 | 34.6 | 34.1 | 28.4 | 34.3 | Trails | Vendor test | Claude Code | Vendor Test | K3/Fable/Sol averaged three official Harbor runs; Fable can fall back to Opus. |
| Coding | MLS Bench Lite | score | 48.3 | 49.9 | 46.2 | 42.8 | 35.5 | 40.4 | Trails | Vendor test | KimiCode | Vendor Test | Claude models use Claude Code and GPT models use Codex. |
| Coding | Kimi Code Bench 2.0 | internal score | 72.9 | 76.9 | 64.8 | 71.7 | 69 | 64.2 | Trails | Internal suite | KimiCode and Claude Code | Vendor Test | In-house Kimi benchmark; all max except GPT-5.5 xhigh. |
| Agentic | GDPval-AA v2 | Elo | 1668 | 1760 | 1748 | 1600 | 1494 | 1514 | Trails | Vendor test | Not specified | Vendor Test | Artificial Analysis source; snapshot can change. |
| Agentic | BrowseComp | score | 91.2 | 88 | 90.4 | 84.3 | 84.4 | Not reported | Leads | Vendor test | Not specified | Vendor Test | Context compaction at 300K; K3 scores 90.4 with raw 1M/no context management. |
| Agentic | DeepSearchQA | F1 | 95 | 94.2 | Not reported | 93.1 | Not reported | Not reported | Leads | Vendor test | Not specified | Vendor Test | Sparse comparator coverage. |
| Agentic | Toolathlon-Verified | score | 73.2 | 77.9 | 74.9 | 76.2 | 73.5 | 59.9 | Trails | Vendor test | Not specified | Vendor Test | Harness details not expanded in launch footnotes. |
| Agentic | MCP Atlas | score | 84.2 | 84.7 | 83.6 | 83.6 | 82.8 | 82.6 | Trails | Vendor test | Not specified | Vendor Test | 500-task public subset; 100-turn limit; Gemini 3.1 Pro judge. |
| Agentic | Automation Bench | score | 30.8 | 29.1 | 29.7 | 27.2 | 22.7 | 12.9 | Leads | Vendor test | Official GitHub setup | Vendor Test | 600-task public subset. |
| Agentic | Job Bench | score | 52.9 | 57.4 | 46.5 | 48.4 | 38.3 | 43.4 | Trails | Vendor test | Not specified | Vendor Test | Vendor table; audit exact benchmark protocol before strong claims. |
| Agentic | AA-Briefcase | Elo | 1548 | 1583 | 1495 | 1354 | 1158 | 1260 | Trails | Vendor test | Not specified | Vendor Test | Artificial Analysis source; snapshot can change. |
| Agentic | APEX-Agents | score | 37.6 | 43.3 | 39.9 | 39.4 | 38.5 | 35.6 | Trails | Vendor test | Not specified | Vendor Test | Vendor table. |
| Agentic | Office QA Pro | score | 63.3 | 69.9 | 63.2 | 63.9 | 60.9 | 41.4 | Trails | Vendor test | Claude Code | Vendor Test | PDFs rendered as images; GPT models use Codex; starred non-K3 scores in source. |
| Agentic | SpreadsheetBench 2 | score | 34.8 | 34.7 | 32.4 | 31.6 | 29.1 | 28.1 | Leads | Vendor test | Claude Code | Vendor Test | GPT models use Codex; starred non-K3 scores in source. |
| Agentic | DECK-Bench | internal score | 73.5 | 73 | 74.7 | 66.9 | 68.2 | 68.6 | Trails | Internal suite | Not specified | Vendor Test | In-house benchmark. |
| Reasoning | GPQA-Diamond | score | 93.5 | 92.6 | 94.1 | 91 | 93.5 | 91.2 | Trails | Vendor test | Not specified | Vendor Test | Vendor table. |
| Reasoning | HLE-Full | score | 43.5 | 53.3 | 44.5 | 49.8 | 41.4 | Not reported | Trails | Vendor test | Not specified | Vendor Test | Opus/GPT-5.5 starred in source; Fable leads this row. |
| Reasoning | HLE-Full with tools | score | 56 | 63 | 58 | 57.9 | 52.2 | Not reported | Trails | Vendor test | Not specified | Vendor Test | Opus/GPT-5.5 starred in source; Fable leads this row. |
| Vision | MMMU-Pro | score | 81.6 | 81.2 | 83 | 78.9 | 81.2 | Not reported | Trails | Vendor test | Official protocol | Vendor Test | Images precede text; most vision rows averaged three runs. |
| Vision | MMMU-Pro with Python | score | 83.4 | 86.5 | 84.6 | 82.7 | 83.2 | Not reported | Trails | Vendor test | Official protocol | Vendor Test | Most vision rows averaged three runs. |
| Vision | CharXiv RQ | score | 84.8 | 88.9 | 84.6 | 80.5 | 84.1 | Not reported | Trails | Vendor test | Not specified | Vendor Test | Most vision rows averaged three runs. |
| Vision | CharXiv RQ with Python | score | 91.3 | 93.5 | 89.1 | 89.9 | 89 | Not reported | Trails | Vendor test | Not specified | Vendor Test | Most vision rows averaged three runs. |
| Vision | MathVision | score | 94.3 | 94.8 | 95.8 | 86.7 | 92.2 | Not reported | Trails | Vendor test | Not specified | Vendor Test | Most vision rows averaged three runs. |
| Vision | MathVision with Python | score | 97.8 | 98.6 | 97.8 | 97.1 | 96.8 | Not reported | Trails | Vendor test | Not specified | Vendor Test | Most vision rows averaged three runs. |
| Vision | BabyVision with Python | score | 85.7 | 90.5 | 88.9 | 81.2 | 83.6 | Not reported | Trails | Vendor test | Not specified | Vendor Test | Most vision rows averaged three runs. |
| Vision | ZeroBench main | pass@5 | 23 | 23 | 17 | 17 | 22 | Not reported | Tied | Vendor test | Official setting | Vendor Test | Run five times; K3 ties Fable. |
| Vision | ZeroBench main with Python | pass@5 | 41 | 46 | 35 | 34 | 41 | Not reported | Trails | Vendor test | Official setting | Vendor Test | Run five times. |
| Vision | WorldVQA ForceAnswer | score | 51 | 56.7 | 41.8 | 39.1 | 38.5 | Not reported | Trails | Vendor test | Not specified | Vendor Test | Most vision rows averaged three runs. |
| Vision | OmniDocBench | score | 91.1 | 89.8 | 85.8 | 87.9 | 89.4 | Not reported | Leads | Vendor test | Not specified | Vendor Test | K3 leads the published row; most vision rows averaged three runs. |
| Vision | PerceptionBench | score | 58.5 | 57.2 | 59.7 | 47.2 | 55.8 | Not reported | Trails | Internal suite | Not specified | Vendor Test | Moonshot in-house atomic-visual-perception benchmark. |
| Independent | Artificial Analysis Intelligence Index v4.1 | index | 57 | 60 | 59 | Not reported | Not reported | Not reported | Trails | Independent | Kimi API | Independent | Independent composite; K3 ranked fourth at capture behind Fable and two Sol effort variants. |
| Independent | Arena WebDev Overall | Elo | 1679 +/-17 (Preliminary) | 1631 | 1618 | Not reported | Not reported | 1587 | Leads | Independent | Arena system | Independent | Preliminary K3 score with +/-17 interval; page captured July 16. |

### Six Cost Presets

| Scenario | Cached input / turn | Cache-miss input / turn | Output / turn | Turns | Worksheet total | Assumptions |
| --- | ---: | ---: | ---: | ---: | ---: | --- |
| Short chat | 0 | 2000 | 1000 | 1 | $0.0210 | One turn; no tools or retries. |
| Coding turn with warm repository | 90000 | 10000 | 20000 | 1 | $0.3570 | 90% of 100K input cached; 20K output/reasoning. |
| Research task | 375000 | 125000 | 80000 | 1 | $1.6875 | 75% of 500K input cached; tools and web fees excluded. |
| Full-context miss | 0 | 1000000 | 100000 | 1 | $4.5000 | One 1M-token request; entry-tier rate limits may conflict. |
| Ten-turn agent loop | 80000 | 20000 | 10000 | 10 | $2.3400 | Each turn averages 100K input at 80% cache and 10K output. |
| Artificial Analysis blended MTok | 700000 | 200000 | 100000 | 1 | $2.3100 | 7:2:1 cache/input/output blend; expressed per 1M blended tokens. |

### Architecture and Operating Guidance

K3 uses a sparse mixture-of-experts architecture with 2.8 trillion total parameters and 16 of 896 experts active per token. Moonshot specifies native vision, a one-million-token context window, MXFP4 weights, and MXFP8 activations.

Moonshot warns that removing preserved thinking history or switching models mid-session can destabilize long-horizon quality. It also warns that ambiguous goals may produce unexpected decisions. Production agents should preserve compatible history and constrain tool permissions.

### Ten-Category Showcase Audit

1. **GPU kernel optimization** (Demonstrated): Moonshot reports up to 24-hour autonomous optimization across four GPU-kernel tasks. Reproduction needs: Publish trajectories, budgets, hardware, tolerances, and same-sandbox independent runs.
2. **MiniTriton compiler** (Demonstrated): Moonshot reports K3 building a Triton-like compiler with MLIR, optimization passes, PTX generation, and nanoGPT training. Reproduction needs: Release the repository, tests, unsupported cases, and an independent performance audit.
3. **Procedural 3D open world** (Demonstrated): Moonshot demonstrates a browser-based procedural 3D open world with vision in the loop. Reproduction needs: Publish the full prompt, code, interventions, asset rights, token budget, and repeat runs.
4. **48-hour chip-design proof of concept** (Demonstrated): Moonshot reports a 48-hour chip-design proof of concept using open-source EDA tools. Reproduction needs: Release RTL, netlists, verification artifacts, PPA methodology, and independent physical validation.
5. **I-Love-Q astrophysics research** (Demonstrated): Moonshot demonstrates an astrophysics workflow spanning papers, equations of state, Python, and an interactive dashboard. Reproduction needs: Publish sources, code, environment, calculation checks, and a domain-expert review.
6. **42-year AI ASIC research site** (Vendor-reported): Moonshot reports a 42-year ASIC-industry research site built through many recursive iterations and source pulls. Reproduction needs: Publish elapsed time, cost, source-quality audit, citation checks, and a human-edit log.
7. **Fusion and gravitational-wave research** (Vendor-reported): Moonshot reports fusion and gravitational-wave research produced with concurrent agents. Reproduction needs: Publish the complete methodology, source set, agent traces, and expert validation.
8. **Slides, Docs, and Sheets artifacts** (Vendor-reported): Moonshot shows editable presentations, documents, spreadsheets, heatmaps, and annual-report artifacts. Reproduction needs: Audit formula accuracy, export fidelity, edit burden, accessibility, and repeatability.
9. **Widgets and Dashboard** (Vendor-reported): Moonshot presents persistent Widgets and Dashboard features connected to data and plugins. Reproduction needs: Document connector coverage, refresh behavior, permissions, data handling, and failure states.
10. **Motion graphics and 56-clip teaser edit** (Demonstrated): Moonshot reports motion graphics and a teaser edit assembled from 56 clips with revisions and beat sync. Reproduction needs: Publish the project file, timeline, human direction, media-rights audit, and a repeat run.

Vendor examples define useful reproduction targets; they do not become independent proof without prompts, budgets, artifacts, intervention logs, and repeat runs.

### Original Test Boundary

K3 Max was selectable anonymously, but the first prompt opened a login modal and returned no output. This page does not claim an independent Matt Farmer K3 model-quality result.

- **T01 / frontend_generation:** not_run_access_blocked
- **T02 / visual_to_code:** not_run_access_blocked
- **T03 / multi_file_debugging:** not_run_access_blocked
- **T04 / long_horizon_recovery:** not_run_access_blocked
- **T05 / cited_research:** not_run_access_blocked
- **T06 / long_context_retrieval:** not_run_access_blocked
- **T07 / office_spreadsheet:** not_run_access_blocked
- **T08 / instruction_boundaries:** not_run_access_blocked. K3 Max was selectable anonymously; prompt submission opened login modal and produced no output.
- **T09 / multilingual:** not_run_access_blocked
- **T10 / multimodal:** not_run_access_blocked

### Privacy

This review did not verify zero retention, no-training-by-default, or data residency as standard Kimi API promises. Review the [current Kimi model-use terms](https://platform.kimi.ai/docs/agreement/modeluse) and request written enterprise terms before sending sensitive data.

### Community Questions

Community reaction identifies questions worth testing. It is not automatically product evidence.

- **Frontier watchers / Frontier shock** (Community, positive): Many users interpret the scores as the first open-weight release to approach the closed frontier. [Source](https://www.reddit.com/r/singularity/comments/1uy9e5n/kimi_k3_benchmarks/).
- **Frontier watchers / Frontend leadership** (Independent, positive): Arena listed K3 first on WebDev Overall at 1679 with a preliminary label. [Source](https://arena.ai/leaderboard/code?rankBy=labs).
- **Frontier watchers / Price pressure** (Community, mixed): Some praise half-Sol sticker pricing; others call $3/$15 expensive for a Chinese open model. [Source](https://www.reddit.com/r/singularity/comments/1uy5ip6/kimi_k3_api_pricing/).
- **Working developers / Token hunger** (Independent, negative): Artificial Analysis measured 130M output tokens across its index; users warn sticker price may understate finished-task cost. [Source](https://artificialanalysis.ai/models/kimi-k3).
- **Working developers / Benchmaxxing concern** (Community, negative): Users want real-world reproduction before replacing established subscriptions. [Source](https://www.reddit.com/r/singularity/comments/1uyniez/you_show_me_kimi_k3_is_not_benchmaxxed_i_cancel/).
- **Working developers / Conflicting real-world quality** (Community, mixed): Some prompts impress; others reportedly trail Sol/Fable or even K2.6 on factuality and task quality. [Source](https://www.reddit.com/r/singularity/comments/1uymqkd/does_k3_really_live_up_to_the_hype_real_world/).
- **Working developers / Persistence** (Community, positive_and_negative): A Kimi user reported a four-hour run that eventually completed; it signals persistence and potentially extreme time/token cost. [Source](https://www.reddit.com/r/kimi/comments/1uyj6ql/kimi_k3_actually_very_impressive/).
- **Local-model operators / Self-hosting excitement** (Community, positive): Users value sovereignty, customization, and reduced vendor lock-in. [Source](https://www.reddit.com/r/singularity/comments/1uy5ip6/kimi_k3_api_pricing/).
- **Local-model operators / Self-hosting reality** (Community, negative): Local-model users joke that a 2.8T release is open but not practically runnable at home. [Source](https://www.reddit.com/r/LocalLLaMA/comments/1uy9cft/kimi_k3_benchmarks/).
- **Frontier watchers / US-China framing** (Community, mixed): Launch coverage and forums frame K3 as evidence that the capability gap is shrinking. [Source](https://www.axios.com/2026/07/16/moonshot-kimi-ai-china-model-openai-anthropic).
- **Working developers / Excessive proactivity** (Official, negative): Moonshot warns K3 may make unexpected decisions on ambiguous tasks. [Source](https://www.kimi.com/blog/kimi-k3).
- **Working developers / Thinking-history sensitivity** (Official, negative): Moonshot warns that missing preserved reasoning history or switching models mid-session can destabilize quality. [Source](https://www.kimi.com/blog/kimi-k3).

## Frequently Asked Questions

### What is Kimi K3?

Kimi K3 is Moonshot AI's 2.8-trillion-parameter sparse mixture-of-experts model with native visual understanding, a 1-million-token context window, agentic products, and an API.

### When did Kimi K3 launch?

Moonshot launched Kimi K3 on July 16, 2026. This evidence snapshot was last verified on July 17, 2026.

### Is Kimi K3 better than GPT-5.6 Sol or Claude Fable 5?

Not overall based on the July 17 evidence. K3 led selected vendor rows and a Preliminary Arena WebDev snapshot, while Artificial Analysis placed Claude Fable 5 and GPT-5.6 Sol ahead on its independent composite.

### How much does the Kimi K3 API cost?

Moonshot lists $0.30 per million cached input tokens, $3 per million cache-miss input tokens, and $15 per million output tokens. Tool fees, retries, and taxes can add to the completed-job cost.

### Is Kimi K3 open source or open weight?

Neither status was verifiable at the July 17 cutoff. Moonshot promised full weights and a technical report by July 27, but the checkpoint and license were not yet available. Open weight and open source are not interchangeable.

### Can I download Kimi K3 now?

No verified full Kimi K3 checkpoint was available on July 17. Use Moonshot's official Hugging Face and GitHub organizations to verify any later release.

### Can Kimi K3 run on a consumer PC or Mac?

Not the full 2.8T model in any practical sense. Moonshot recommends a supernode with 64 or more accelerators for deployment; that is vendor guidance, not an independently measured minimum.

### Why does preserved thinking history matter?

Moonshot says K3 was trained with preserved thinking history and can become unstable when a harness omits the complete historical assistant message or switches models mid-session.

### What does Moonshot mean by excessive proactivity?

Moonshot warns that K3 may make unexpected decisions when a task has minor problems or ambiguous intent. Production agents should use explicit permissions, budgets, stop conditions, and approval boundaries.

### What did independent Kimi K3 benchmarks find?

At the July 17 capture, Artificial Analysis scored K3 at 57 with 62 output tokens per second and 130 million output tokens across its Intelligence Index run. Arena listed K3 first on WebDev Overall at 1679 +/-17, marked Preliminary, on July 16.

### Why can a cheaper per-token model cost more per task?

Completed-job cost depends on output volume, turns, retries, caching, tool fees, latency, and cleanup. K3's $15-per-million output rate can dominate the bill when the model reasons or retries at length.

### Does Moonshot use API data for model improvement?

Moonshot's captured API terms say customer content may be used to provide, maintain, develop, support, improve, secure, and enforce the service unless a separate written enterprise arrangement restricts that use. Review the current terms before sending sensitive data.

### What changes on July 27, 2026?

Moonshot promised full model weights by July 27. A responsible update must verify the checkpoint, license, technical report, model configuration, vLLM support, third-party hosts, and independent self-host evidence before changing the page's pending status.

## Research Changelog

- **v0.1, Jul 15, 2026:** Pre-launch rumor-control packet kept the model ID, price, size, context, and access claims unverified.
- **v1.0, Jul 17, 2026:** Post-launch packet added official specifications, pricing, benchmark rows, independent measurements, access status, terms, and the July 27 watchlist.
- **v1.0 check, Jul 17, 2026:** Publication check confirmed the launch, pricing, Arena Preliminary result, Artificial Analysis snapshot, terms, and that K3 weights/report were still absent from Moonshot's public organizations.

## Complete Source Index

1. [Moonshot: Kimi K3 launch](https://www.kimi.com/blog/kimi-k3) - Specifications, benchmark table, availability, architecture, deployment guidance, and limitations
2. [Kimi K3 API quickstart](https://platform.kimi.ai/docs/guide/kimi-k3-quickstart) - Model ID, context, API behavior, max effort, multimodal input, caching, and limits
3. [Kimi K3 API pricing](https://platform.kimi.ai/docs/pricing/chat-k3) - Official pricing and billing notes
4. [Kimi API rate limits](https://platform.kimi.ai/docs/pricing/limits) - Recharge and rate-limit terms
5. [Kimi OpenPlatform terms](https://platform.kimi.ai/docs/agreement/modeluse) - Customer-content use and enterprise agreement language
6. [Kimi Code model configuration](https://www.kimi.com/code/docs/en/kimi-code/models.html) - Kimi Code plan access, context, and low/high/max effort listing
7. [Moonshot AI on Hugging Face](https://huggingface.co/moonshotai) - Negative checkpoint check at the July 17 cutoff
8. [Moonshot AI on GitHub](https://github.com/MoonshotAI) - Negative technical-report and repository check at the July 17 cutoff
9. [Artificial Analysis: Kimi K3](https://artificialanalysis.ai/models/kimi-k3) - Independent intelligence, speed, latency, pricing, and token-use snapshot
10. [Arena WebDev leaderboard](https://arena.ai/leaderboard/code?rankBy=labs) - Independent Preliminary WebDev Overall snapshot
11. [Frontier watchers: Frontier shock](https://www.reddit.com/r/singularity/comments/1uy9e5n/kimi_k3_benchmarks/) - Anecdotal community evidence checked 2026-07-17
12. [Frontier watchers: Frontend leadership](https://arena.ai/leaderboard/code?rankBy=labs) - Independent evidence checked 2026-07-17
13. [Local-model operators: Self-hosting excitement](https://www.reddit.com/r/singularity/comments/1uy5ip6/kimi_k3_api_pricing/) - Anecdotal community evidence checked 2026-07-17
14. [Working developers: Token hunger](https://artificialanalysis.ai/models/kimi-k3) - Independent evidence checked 2026-07-17
15. [Working developers: Benchmaxxing concern](https://www.reddit.com/r/singularity/comments/1uyniez/you_show_me_kimi_k3_is_not_benchmaxxed_i_cancel/) - Anecdotal community evidence checked 2026-07-17
16. [Working developers: Conflicting real-world quality](https://www.reddit.com/r/singularity/comments/1uymqkd/does_k3_really_live_up_to_the_hype_real_world/) - Anecdotal community evidence checked 2026-07-17
17. [Working developers: Persistence](https://www.reddit.com/r/kimi/comments/1uyj6ql/kimi_k3_actually_very_impressive/) - Anecdotal community evidence checked 2026-07-17
18. [Local-model operators: Self-hosting reality](https://www.reddit.com/r/LocalLLaMA/comments/1uy9cft/kimi_k3_benchmarks/) - Anecdotal community evidence checked 2026-07-17
19. [Frontier watchers: US-China framing](https://www.axios.com/2026/07/16/moonshot-kimi-ai-china-model-openai-anthropic) - Anecdotal community evidence checked 2026-07-17
20. [Working developers: Thinking-history sensitivity](https://www.kimi.com/blog/kimi-k3) - Official evidence checked 2026-07-17