DeepSeek V4.1 Flash vs Kimi K3

At a glance

• Capability index

  • DeepSeek V4.1 Flash155.6
  • Kimi K3157.4

  Kimi K3 is 1.8 points higher.

• Price per 1M tokens

  • DeepSeek V4.1 Flash$0.464
  • Kimi K3$6.00

  DeepSeek V4.1 Flash is 13 times cheaper.

• Context window

  • DeepSeek V4.1 Flash1M tokens
  • Kimi K31M tokens

  They take in about the same.

• Released

  • DeepSeek V4.1 FlashSep 10, 2026
  • Kimi K3Jul 16, 2026

  DeepSeek V4.1 Flash is 56 days newer.

Where each scores higher

On the 21 tests both have taken, each on its own scale. A gap of a point or two can sit within a test’s margin of error.

1 more test has results for only one of them; the table below lists every result.

Everything side by side

Comparison of DeepSeek V4.1 Flash vs Kimi K3

Measure	DeepSeek V4.1 FlashDeepSeek	Kimi K3Moonshot AI
Model
Maker	DeepSeek	Moonshot AI
Sold here by	engy	Moonshot AI
API model ID	deepseek-v4.1-flash	kimi-k3
Released	Sep 10, 2026	Jul 16, 2026
Knowledge cutoff	May 2025	Not reported
Weights	Open: downloadable	Open: downloadable
Developer’s country	China	China
Hosts selling it	80	89including Moonshot AI directly
Training compute	4.3 × 10²⁴ FLOPEpoch AI estimate	2.0 × 10²⁵ FLOPEpoch AI estimate
Test results
Capability index	155.6#31 of 277	157.4 (best of these models)#17 of 277
LiveBench overall	81.1% (best of these models)#9 of 70	79.2%#16 of 70
Arena text	1475#38 of 415	1488 (best of these models)#17 of 415
LiveBench coding	80.0%#23 of 70	81.5% (best of these models)#15 of 70
Arena WebDev	1619#24 of 142	1655 (best of these models)#14 of 142
Arena coding	1526#27 of 410	1539 (best of these models)#10 of 410
LiveBench mathematics	93.3% (best of these models)#23 of 70	84.4%#57 of 70
Mock AIME 2024–2025	98.3% (best of these models)#24 of 303	97.2%#36 of 303
FrontierMath Tiers 1–3	67.4%#30 of 117	72.2% (best of these models)#24 of 117
FrontierMath Tier 4	26.8%#41 of 73	39.0% (best of these models)#31 of 73
Arena math	1495#22 of 400	1500 (best of these models)#17 of 400
SimpleQA Verified	42.9%#43 of 90	50.6% (best of these models)#25 of 90
LiveBench agentic coding	77.3% (best of these models)#1 of 70	62.2%#11 of 70
DeepSWE	Not tested	68.5%#19 of 69
LiveBench language	81.2%#29 of 70	85.5% (best of these models)#11 of 70
LiveBench instructions	70.0%#32 of 70	71.4% (best of these models)#24 of 70
Arena creative writing	1439#61 of 413	1460 (best of these models)#29 of 413
Arena instruction following	1477#24 of 415	1488 (best of these models)#15 of 415
LiveBench data analysis	79.3% (best of these models)#18 of 70	78.7%#22 of 70
LiveBench reasoning	86.7%#32 of 70	90.7% (best of these models)#9 of 70
GPQA Diamond	89.8%#50 of 323	93.1% (best of these models)#21 of 323
Chess puzzles	48.0% (best of these models)#20 of 243	37.0%#46 of 243
Arena hard prompts	1499#36 of 415	1517 (best of these models)#11 of 415
Price per million tokens
Input	$0.04 (best of these models)	$3.00
Output	$0.08 (best of these models)	$15.00
Blended, 3 input to 1 output	$0.05 (best of these models)	$6.00
Cached input	$0.008 (best of these models)	$0.30
Writing to the cache	Not reported	$3.00
The maker’s own price	Not sold directly	This listing
Limits
Context window	327.7K tokens	1M tokens (best of these models)
Max input	1M tokens	1M tokens
Max output	65.5K tokens	1M tokens (best of these models)
Features
Reads	images and text	images, text and video
Produces	text	text
Reasoning	Yeseffort none, low, high, max; can be switched off; thinking budget can be set	Yeseffort low, high, max; can be switched off; thinking budget can be set
Tool calling	Yes	Yes
Structured output	Yes, such as JSON	Yes, such as JSON
File attachments	Yes	Yes
Temperature setting	Supported	Fixed by the provider

Bold marks the better value in each row. Prices are each model’s own list price, or the middle price across hosts when the maker does not sell it, blended as three input tokens for every output token. Test results as published by LiveBench (Apache-2.0), Epoch AI (CC BY 4.0), Datacurve DeepSWE, compiled by Epoch AI (CC BY 4.0), Arena (CC BY 4.0); model details and prices from models.dev (MIT). Confirm pricing with the provider before use.