Model comparison
Kimi vs DeepSeek: Kimi K3 and DeepSeek V4.1 Flash
Choose DeepSeek V4.1 Flash for the higher reported DeepSWE v1.1 score (74.2% vs 67.5%); DeepSeek V4.1 Flash has the lower fixed-token cost, $0.00420 for 10,000 input and 1,000 billed output tokens.
| Dimension | Winner | Margin | Evidence |
|---|---|---|---|
| Coding benchmark | DeepSeek V4.1 Flash | +6.7 pp reported score gap | DeepSWE v1.1: DeepSeek V4.1 Flash 74.2% vs Kimi K3 67.5%1 |
| Reasoning benchmark | Kimi K3 | +2.0 pp reported score gap | GPQA Diamond: Kimi K3 92.9% vs DeepSeek V4.1 Flash 90.9%1 |
| Estimated request cost | DeepSeek V4.1 Flash | $0.04080 difference per example | DeepSeek V4.1 Flash $0.00420 vs Kimi K3 $0.0450023 |
Official benchmark results
Official benchmark results
Swipe to see all columns
| Metric | Kimi K3 | DeepSeek V4.1 Flash |
|---|---|---|
| DeepSWE v1.1 (Resolved)1 | 67.5% · vendor-reported, 2026 | 74.2% · vendor-reported, 2026 |
| GPQA Diamond (Pass@1)1 | 92.9% · vendor-reported, 2026 | 90.9% · vendor-reported, 2026 |
| HLE w/ tools (Pass@1)1 | 59.8% · vendor-reported, 2026 | 63.9% · vendor-reported, 2026 |
| NL2Repo-Bench1 | 58.0 · vendor-reported, 2026 | 64.0 · vendor-reported, 2026 |
| HLE (Pass@1)1 | 43.5% · vendor-reported, 2026 | 36.8% · vendor-reported, 2026 |
Performance
Performance
Swipe to see all columns
Pricing
Pricing
Swipe to see all columns
| Metric | Kimi K3 | DeepSeek V4.1 Flash |
|---|---|---|
| Input / 1M tokens23 | $3.00 | $0.30 |
| Output / 1M tokens23 | $15.00 | $1.20 |
| Context window (tokens)23 | 1M | 1M |
| Max output (tokens)23 | 1M | 384K |
| API and model ID23 | moonshotai/kimi-k3 | deepseek/deepseek-v4.1-flash |
Which to choose
Coding, reasoning, context and cost
Kimi K3
- Prefer Kimi K3 for the higher GPQA and HLE-without-tools report scores: 92.9% and 43.5%, versus DeepSeek’s 90.9% and 36.8%.
- For a larger completion budget, Kimi lists 1M max output versus 384K; both input windows are 1M.
- The trade-off is a $0.04500 request estimate versus $0.00420, and lower DeepSWE at 67.5% versus 74.2%.
DeepSeek V4.1 Flash
- Prefer DeepSeek for repository agents: 74.2% DeepSWE and 64.0 NL2Repo, versus Kimi’s 67.5% and 58.0.
- At $0.00420 per fixed-token example, DeepSeek saves $0.04080 relative to Kimi; generated reasoning remains billable.
- Its 1M context supports large inputs, but 384K max output is below Kimi’s 1M completion cap.
Methodology
Methodology and sources
The Kimi K3 and DeepSeek V4.1 Flash benchmark rows share DeepSeek's published maximum-effort comparison. HLE results retain their with-tools or without-tools conditions. DeepSeek's Novita endpoint uses fp8 and defaults to high; its official and AA Max references are distinct from that default. AA measures reference-model streaming speed rather than a paired Novita endpoint test.
This Kimi and DeepSeek page is published by Novita AI using public evidence. Its USD example applies uncached synchronous prices to 10,000 input and 1,000 billed output tokens, with billed reasoning included. Tax, retries, discounts and dedicated capacity are outside that estimate; it is not an invoice or a private workload result.
Kimi and DeepSeek first-answer estimates assume 2,000 reasoning tokens. The calculation uses each model's unrounded AA streaming rate and first-chunk delay; rounded table entries are for display, and the illustrative budget does not predict a particular prompt's thinking time.
- DeepSeek · V4.1 Flash Model Card huggingface.co · Checked 2026-10-15
- Kimi K3 · Novita AI novita.ai · Checked 2026-10-15
- DeepSeek V4.1 Flash · Novita AI novita.ai · Checked 2026-10-15
- Kimi K3 · Artificial Analysis artificialanalysis.ai · Checked 2026-10-15
- DeepSeek V4.1 Flash · Artificial Analysis artificialanalysis.ai · Checked 2026-10-15
Build with Novita Model APIs
Create an API key and connect your application.