Instruction following
Like-for-like- DeepSeek V3
- 86.1
- Kimi K2.5
- 93.9
- Weighted basis
- 1 vs 1 rows
- Reading
- Kimi K2.5 leads
Model comparison
DeepSeek V3 vs Kimi K2.5
Updated July 30, 2026. Public scores include evidence status and uncertainty. They are not guarantees for a specific workload.
Kimi K2.5 has the higher public score estimate, 58.78 versus 44.15, but the 90% score intervals overlap. Treat that as a lead, not a settled winner.
5 results are shared. Category rows based on different benchmark sets are marked directional and do not name a winner.
Which one for your work
Recommendations appear only when a shared evidence basis or an explicit operating constraint supports the call. Secondary and unsupported use cases stay disclosed below the initial list.
Long documents
Prompts that approach the documented context limit
Kimi K2.5
Kimi K2.5 has the larger documented context window.
Confidence: documented
Chat turn cost
1K fresh input + 500 output tokens
DeepSeek V3
DeepSeek V3 has the lower estimated token cost for this stated workload. Costs use the listed standard API rates.
Confidence: listed-rates
Show secondary and unsupported calls
Repository review cost
50K fresh input + 3K output tokens
DeepSeek V3
DeepSeek V3 has the lower estimated token cost for this stated workload. Costs use the listed standard API rates.
Confidence: listed-rates
Coding work
Code generation, repair, and software-engineering tasks
Not enough matched evidence
The category averages use different weighted benchmark sets, so they are directional rather than like-for-like.
Confidence: limited
Agentic work
Tool use, computer use, and multi-step task completion
Not enough matched evidence
No shared weighted benchmark basis supports a winner.
Confidence: limited
Cache-heavy agent loop cost
200K cached + 20K fresh input + 10K output tokens
Not enough matched evidence
The page does not recommend a cost winner because at least one model cannot fit the stated workload in one request. DeepSeek V3 does not fit this workload in one request. Kimi K2.5 has no published cached-input rate, so cached tokens use its listed input rate.
Confidence: rate-fallback
What is actually comparable
Shared results can support a head-to-head reading. Results present for only one model describe coverage, not superiority.
1 DeepSeek V3 only
5 Shared
40 Kimi K2.5 only
- Shared results
- 5
- DeepSeek V3 only
- 1
- Kimi K2.5 only
- 40
- Like-for-like categories
- 1 / 8
3 categories use different evidence sets. Those rows remain visible for coverage context but do not name a winner.
Category results, on a stated basis
Each row states whether both averages use the same weighted benchmark set. Directional and not-comparable rows remain visible, but they never receive a winner in this template.
Coding
Directional only- DeepSeek V3
- 38.9
- Kimi K2.5
- 59.4
- Weighted basis
- 2 vs 4 rows
- Reading
- Directional only
Knowledge
Directional only- DeepSeek V3
- 72.7
- Kimi K2.5
- 56.9
- Weighted basis
- 2 vs 4 rows
- Reading
- Directional only
Math
Directional only- DeepSeek V3
- 1.7
- Kimi K2.5
- 60.6
- Weighted basis
- 1 vs 4 rows
- Reading
- Directional only
Agentic
Not comparable- DeepSeek V3
- Not measured
- Kimi K2.5
- 55.0
- Weighted basis
- 0 vs 2 rows
- Reading
- Not comparable
Reasoning
Not comparable- DeepSeek V3
- Not measured
- Kimi K2.5
- 61.0
- Weighted basis
- 0 vs 1 rows
- Reading
- Not comparable
Multilingual
Not comparable- DeepSeek V3
- Not measured
- Kimi K2.5
- 82.3
- Weighted basis
- 0 vs 1 rows
- Reading
- Not comparable
Multimodal
Not comparable- DeepSeek V3
- Not measured
- Kimi K2.5
- 78.5
- Weighted basis
- 0 vs 1 rows
- Reading
- Not comparable
| Category | DeepSeek V3 | Kimi K2.5 | Weighted basis | Reading |
|---|---|---|---|---|
| Instruction following | 86.1 | 93.9 | Like-for-like1 vs 1 rows | Kimi K2.5 leads |
| Coding | 38.9 | 59.4 | Directional only2 vs 4 rows | Directional only |
| Knowledge | 72.7 | 56.9 | Directional only2 vs 4 rows | Directional only |
| Math | 1.7 | 60.6 | Directional only1 vs 4 rows | Directional only |
| Agentic | Not measured | 55.0 | Not comparable0 vs 2 rows | Not comparable |
| Reasoning | Not measured | 61.0 | Not comparable0 vs 1 rows | Not comparable |
| Multilingual | Not measured | 82.3 | Not comparable0 vs 1 rows | Not comparable |
| Multimodal | Not measured | 78.5 | Not comparable0 vs 1 rows | Not comparable |
Shape of the matched evidence
Only shared public evidence is shown. Sparse evidence stays a ruled list rather than being closed into a radar shape.
Too few matched category axes support a radar. The ruled list below shows only shared benchmark results; positions use each benchmark’s normalized display scale when available.
SWE-bench Verified
Coding
GPQA
Knowledge
FrontierMath v2 (Tiers 1-3)
Math
MMLU-Pro
Knowledge
IFEval
Instruction following
What each workload costs
Three fixed token mixes turn per-token rates into comparable decisions. Each scenario states context fit and whether cached input had to fall back to the published list-input rate.
Chat turn
1K fresh input + 500 output tokens
- DeepSeek V3
- $0.00082
- Fits in one request
- Kimi K2.5
- $0.0021
- Fits in one request
DeepSeek V3 has the lower modeled cost
Costs use the listed standard API rates.
Repository review
50K fresh input + 3K output tokens
- DeepSeek V3
- $0.0168
- Fits in one request
- Kimi K2.5
- $0.039
- Fits in one request
DeepSeek V3 has the lower modeled cost
Costs use the listed standard API rates.
Cache-heavy agent loop
200K cached + 20K fresh input + 10K output tokens
- DeepSeek V3
- $0.0304
- Does not fit in one request
- Kimi K2.5
- $0.162
- Fits in one request
- Cached input priced at the published list-input rate
DeepSeek V3 does not fit this workload in one request. Kimi K2.5 has no published cached-input rate, so cached tokens use its listed input rate.
Specification differences
Sourced differences are shown directly. Missing facts stay explicit instead of being inferred from a model name or family.
SpecificationDeepSeek V3Kimi K2.5
Context window
Maximum documented context; output-token limits may be lower.
DeepSeek V3
128K
Kimi K2.5
256K
API model ID
DeepSeek V3
Not sourced
Kimi K2.5
Not sourced
Cached-input rate
A missing cached-input rate falls back to the listed input rate only in the stated workload estimate.
DeepSeek V3
$0.07 per 1M cached input tokens
Kimi K2.5
Not published
Documented inputs
DeepSeek V3
Not sourced
Kimi K2.5
Not sourced
Documented outputs
DeepSeek V3
Not sourced
Kimi K2.5
Not sourced
Provider availability
DeepSeek V3
Not sourced
Kimi K2.5
Not sourced
Reasoning profile
DeepSeek V3
Non-Reasoning
Kimi K2.5
Non-Reasoning
Weight access
DeepSeek V3
Open Weight
Kimi K2.5
Open Weight
License
DeepSeek V3
Open Weight
Kimi K2.5
Open Weight
Release date
DeepSeek V3
2024-12-26
Kimi K2.5
2026-02-01
If you already use one of these models
- Deployment change
- The models list different providers, so authentication, endpoint behavior, limits, and feature support may change.
- Quality signal
- Kimi K2.5 has the higher public score estimate, 58.78 versus 44.15, but the 90% score intervals overlap.
- Workload cost
- Repository review: $0.0168 vs $0.039. Cache-heavy agent loop: $0.0304 vs $0.162.
- Context tradeoff
- Kimi K2.5 has the larger documented window (256K).
Run the same representative tasks against both endpoints before changing production traffic.
Self-host vs API cost
Estimates at 50,000 req/day · 1000 tokens/req average.
DeepSeek V3
API / mo$1,028
Self-host / mo$18,221
Break-even1.2B/day
Kimi K2.5
API / mo$2,700
Self-host / mo$5,221
Break-even132M/day
Benchmark evidence
The full public result ledger is available for audit without forcing a wide desktop table onto a phone.
Browse raw public benchmark evidence46 rows
Agentic
Terminal-Bench 2.0
DeepSeek V3—Kimi K2.550.8%SourceNot directly comparable
BrowseComp
DeepSeek V3—Kimi K2.560.6%SourceNot directly comparable
Claw-Eval
DeepSeek V3—Kimi K2.552.3%SourceNot directly comparable
QwenClawBench
DeepSeek V3—Kimi K2.554.3%SourceNot directly comparable
τ³-bench results
DeepSeek V3—Kimi K2.565.7%SourceNot directly comparable
DeepSearchQA
DeepSeek V3—Kimi K2.577.1%SourceNot directly comparable
DeepPlanning
DeepSeek V3—Kimi K2.514.4%SourceNot directly comparable
Toolathlon
DeepSeek V3—Kimi K2.527.8%SourceNot directly comparable
MCP Atlas
DeepSeek V3—Kimi K2.529.5%SourceNot directly comparable
MCP-Tasks
DeepSeek V3—Kimi K2.559.1%SourceNot directly comparable
WideResearch
DeepSeek V3—Kimi K2.572.7%SourceNot directly comparable
Gert Labs
DeepSeek V3—Kimi K2.545.88%SourceNot directly comparable
ResearchClawBench
DeepSeek V3—Kimi K2.514.0%SourceNot directly comparable
JobBench
DeepSeek V3—Kimi K2.58.7%SourceNot directly comparable
Coding
LiveCodeBench
DeepSeek V337.6%SourceKimi K2.5—Not directly comparable
SWE-bench Verified
DeepSeek V342%SourceKimi K2.576.8%SourceKimi K2.5 leads this result
SWE-bench Verified*
DeepSeek V3—Kimi K2.570.8%SourceNot directly comparable
LiveCodeBench v6
DeepSeek V3—Kimi K2.585.0%SourceNot directly comparable
SWE-bench Pro
DeepSeek V3—Kimi K2.550.7%SourceNot directly comparable
SWE Multilingual
DeepSeek V3—Kimi K2.573%SourceNot directly comparable
SWE-Rebench
DeepSeek V3—Kimi K2.558.5%SourceNot directly comparable
React Native Evals
DeepSeek V3—Kimi K2.577.2%SourceNot directly comparable
SciCode
DeepSeek V3—Kimi K2.548.7%SourceNot directly comparable
Reasoning
LongBench v2
DeepSeek V3—Kimi K2.561%SourceNot directly comparable
Knowledge
GPQA
DeepSeek V359.1%SourceKimi K2.587.6%SourceKimi K2.5 leads this result
MMLU-Pro
DeepSeek V375.9%SourceKimi K2.587.1%SourceKimi K2.5 leads this result
GPQA-D
DeepSeek V3—Kimi K2.587.6%SourceNot directly comparable
SuperGPQA
DeepSeek V3—Kimi K2.569.2%SourceNot directly comparable
MMLU-Pro (Arcee)
DeepSeek V3—Kimi K2.587.1%SourceNot directly comparable
HLE
DeepSeek V3—Kimi K2.530.1%SourceNot directly comparable
Math
FrontierMath v2 (Tiers 1-3)
Shared sourceDeepSeek V31.724%Kimi K2.527.900%Kimi K2.5 leads this result
AIME 2025
DeepSeek V3—Kimi K2.596.1%SourceNot directly comparable
AIME26
DeepSeek V3—Kimi K2.595.8%SourceNot directly comparable
AIME25 (Arcee)
DeepSeek V3—Kimi K2.596.3%SourceNot directly comparable
HMMT Feb 2025
DeepSeek V3—Kimi K2.595.4%SourceNot directly comparable
HMMT Nov 2025
DeepSeek V3—Kimi K2.591.1%SourceNot directly comparable
HMMT Feb 2026
DeepSeek V3—Kimi K2.587.1%SourceNot directly comparable
MMAnswerBench
DeepSeek V3—Kimi K2.581.8%SourceNot directly comparable
FrontierMath v2 (Tier 4)
DeepSeek V3—Kimi K2.54.200%SourceNot directly comparable
Multilingual
Multimodal
Instruction following
Frequently asked questions
Which is better, DeepSeek V3 or Kimi K2.5?
Kimi K2.5 has the higher public score estimate, 58.78 versus 44.15, but the 90% score intervals overlap. The higher estimate is not a decisive winner because the uncertainty ranges overlap.
Which is better for coding, DeepSeek V3 or Kimi K2.5?
The current coding averages use different weighted benchmark sets, so BenchLM does not name a winner from them. Read the shared benchmark rows directly and test the models on the same task set.
Which is better for agentic tasks, DeepSeek V3 or Kimi K2.5?
The published evidence does not provide a shared weighted agentic tasks basis for both models, so BenchLM does not name a agentic tasks winner.
Which costs less, DeepSeek V3 or Kimi K2.5?
For the stated presets, chat costs $0.00082 on DeepSeek V3 and $0.0021 on Kimi K2.5; repository review costs $0.0168 and $0.039; the cache-heavy agent loop costs $0.0304 and $0.162. DeepSeek V3 does not fit this workload in one request. Kimi K2.5 has no published cached-input rate, so cached tokens use its listed input rate.
Which has the larger context window, DeepSeek V3 or Kimi K2.5?
Kimi K2.5 has the larger documented context window: 256K, compared with 128K.
Related comparisons
Last updated July 30, 2026
Watch DeepSeek V3 vs Kimi K2.5
One weekly email when material rank, price, or benchmark evidence changes make this matchup worth revisiting.
Read a sample issueJoin 2,000+ readers.