none reviewed yet · 4 benchmark observations · latest capture Sep 8, 2026
Inspect the evidenceMoonshot AI · MODEL PROFILE
Kimi K3
Design-led teams and self-hosters who want the best visual-preference model with open weights
Saved editorial fit for your current task lens; not a benchmark result
How scoring worksVerified rates (Sep 8, 2026) · 100 runs / month at your token settings
Adjust workloadAt a glance
9 facts- Provider
- Moonshot AI
- Type
- API model
- Access
- platform.kimi.ai (OpenAI- and Anthropic-compatible), Hugging Face weights, OpenRouter (15 providers), Alibaba Cloud Model Studio; Kimi Code CLI is the recommended agent harness
- Context
- 1,048,576 tokens
- Maximum output
- OpenRouter lists up to 1,048,576 completion tokens; the model card does not state a max output
- License / weights
- Open weights
- Input modalities
- text, image, video
- Output modalities
- text
- Pricing notes
- $3 in / $15 out per 1M with $0.30 cached input (Moonshot); Alibaba resells at $2.827/$14.133 in most regions and $3/$15 in Singapore/US international; OpenRouter median $2.55/$12.75 with $0.256 cache read
The practical trade-offs
Documented strengths
- Design Arena Website rank 1 (1361) - the best pure design-preference score in this set
- Arena WebDev Frontend rank 5 (1691) - top open-weight model on that board after Qwen3.8 Max
- Open weights at 2.8T total / 104B active with a 1M context and native vision
- MCPMark-Verified 94.5 and Toolathlon 76.5 make it a genuinely strong tool user
Limitations to check
- Self-hosting means a ~1.56 TB download and multi-node serving
- $3/$15 API pricing is high for an open-weight model
- AA measured only 40.3 tok/s (OpenRouter providers reach 90)
- AA index figures in circulation conflict (50 at max vs 57 claimed at launch)
Follow the source
8 source linksEvidence & conditions
8Source links
NoneReviewed records
4Benchmark runs
Sep 8, 2026Latest capture
- Evidence source 1
huggingface.coDate not recorded - Evidence source 2
k3-kimi.comDate not recorded - Evidence source 3
benchlm.aiDate not recorded - Evidence source 4
arena.aiDate not recorded - Evidence source 5
benchlm.aiDate not recorded - Evidence source 6
artificialanalysis.aiDate not recorded - Evidence source 7
openrouter.aiDate not recorded - Evidence source 8
alibabacloud.comDate not recorded
Benchmark observations
Terminal-Bench 2.1 Inspect conditions
- name
- Terminal-Bench 2.1
- value
- 88.3
- unit
- %
- source Url
- https://huggingface.co/moonshotai/Kimi-K3
- checked At
- 2026-09-08T19:28:18Z
- attribution
- provider-reported
- configuration
- Kimi K3; Kimi Code harness; max effort; temperature=1; top_p=1 for agentic evaluation.
DeepSWE v1.1 Inspect conditions
- name
- DeepSWE v1.1
- value
- 67.5
- unit
- %
- source Url
- https://huggingface.co/moonshotai/Kimi-K3
- checked At
- 2026-09-08T19:28:18Z
- attribution
- provider-reported
- configuration
- Kimi K3; Kimi Code harness; max effort; temperature=1; top_p=1 for agentic evaluation. This differs from the separately reported mini-SWE-agent result.
ProgramBench Inspect conditions
- name
- ProgramBench
- value
- 77.8
- unit
- %
- source Url
- https://huggingface.co/moonshotai/Kimi-K3
- checked At
- 2026-09-08T19:28:18Z
- attribution
- provider-reported
- configuration
- Kimi K3; Kimi Code harness; max effort; temperature=1; top_p=1 for agentic evaluation.
Arena WebDev Frontend Inspect conditions
- name
- Arena WebDev Frontend
- value
- 1691
- unit
- arena_score
- source Url
- https://arena.ai/leaderboard/code/webdev/frontend
- checked At
- 2026-09-08T19:28:18Z
- attribution
- benchmark-publisher
- configuration
- Exact row kimi-k3-max; source results dated 2026-09-05; human preference; confidence interval +/-13; 3744 votes. This is a checked source snapshot, not a newly run September 8 evaluation.
Source reports, measured results, and editorial assessments are separate evidence types. Source dates do not imply a new benchmark run. Capture dates are shown when recorded; otherwise we say the capture date was not recorded. Recorded dates are compared with this catalogue snapshot (Sep 11, 2026), not today’s date.
Method AIS-V2
How this page is assessed
Quality claims follow Method AIS-V2.
Money never buys rank — commercial relationships cannot change rank, eligibility, evidence quality, or publication timing.
Submit or claim a listing is free and does not skip the editorial queue.
Private · this device only