Model comparison
Kimi K2.5 vs Kimi K3
Kimi K3 is the stronger model overall, scoring 59.5 to 48.1 on the Noometry Index. Kimi K2.5 costs 6.7× less per token, which makes it the better buy when Kimi K3's lead doesn't matter for your workload.
Last verified . 34 shared benchmarks.
Summary
- They share 34 benchmarks with published results for both. Kimi K2.5 scores higher in 2 categories and Kimi K3 in 8 categories; 10 gaps are clear of the uncertainty.
- The widest gap is in reasoning, where Kimi K3 leads 63.0 to 31.2.
- The biggest single-benchmark swing is ARC-AGI-2: 11.8% for Kimi K2.5 and 60.4% for Kimi K3.
- Kimi K2.5 is cheaper at $0.45 / $2.25 per million input/output tokens, against $3 / $15 for Kimi K3.
- Kimi K3 accepts more context: 1.05M tokens versus 262K.
Side by side
| Kimi K2.5 | Kimi K3 | |
|---|---|---|
| Provider | Moonshot AI | Moonshot AI |
| Noometry Index | 48.1 | 59.5 |
| Released | 2026-01-27 | 2026-07-16 |
| Weights | Open | Open |
| Context window | 262K | 1.05M |
| Max output | 262K | 1.05M |
| Input $ / M tokens | $0.45 | $3 |
| Output $ / M tokens | $2.25 | $15 |
| Results tracked | 51 | 53 |
Sponsored placements are available on pages like this one. Advertise on Noometry
Category by category
Coding Kimi K3 leads
Kimi K2.5: 48.8 (#53), Kimi K3: 61.0 (#10)
| Benchmark | Kimi K2.5 | Kimi K3 |
|---|---|---|
| LMArena WebDev | 1437 | 1654 |
| SciCode | 49% | 59.5% |
| WeirdML | 45.6% | 82.6% |
| LMArena Coding | 1474 | 1508 |
| ALE-Bench | 821.65 | 1,524 |
| SWE-bench Verified | 73.8% | — |
| DeepSWE | — | 68.5% |
| FrontierCode | — | 44.2% |
| SWE-bench Verified (bash only) | 70.8% | — |
| SWE-bench Multilingual | 67.3% | — |
| FrontierSWE | — | 25.9% |
Agentic & Tool Use Kimi K3 leads
Kimi K2.5: 34.2 (#48), Kimi K3: 41.8 (#20)
| Benchmark | Kimi K2.5 | Kimi K3 |
|---|---|---|
| Vending-Bench 2 | 1,198 | 5,165 |
| Terminal-Bench | 43.2% | — |
| APEX-Agents | — | 50.6% |
| τ²-bench Banking | — | 37.1% |
| OSWorld | 63.3% | — |
| PostTrainBench | — | 32% |
| GBAEval | — | 48.3% |
| GDP.pdf | — | 19% |
Reasoning Kimi K3 leads
Kimi K2.5: 31.2 (#80), Kimi K3: 63.0 (#17)
| Benchmark | Kimi K2.5 | Kimi K3 |
|---|---|---|
| ARC-AGI-2 | 11.8% | 60.4% |
| SimpleBench | 46.8% | 60.7% |
| NYT Connections (extended) | 69.9% | 93.6% |
| ARC-AGI-1 | 65.3% | 94.5% |
| CritPt | 3.1% | 23.4% |
| Chess Puzzles | 12% | 39% |
| LMArena Hard Prompts | 1453 | 1496 |
| Epoch Capabilities Index | 148.03 | 157.45 |
| Kagi LLM Benchmark | 78.5% | — |
| EnigmaEval | 3.4% | — |
| Thematic Generalization | 69.4% | — |
| Mystery Game Puzzles | — | 26% |
| DTBench | — | 91.2% |
| LMCA | — | 52.7% |
| Surface Evolver Bench | — | 95% |
| ForecastBench | — | 61.1 |
Math Kimi K3 leads
Kimi K2.5: 51.8 (#53), Kimi K3: 74.2 (#16)
| Benchmark | Kimi K2.5 | Kimi K3 |
|---|---|---|
| MathArena Final-Answer Competitions | 62.3% | 87.8% |
| OTIS Mock AIME 2024-2025 | 92.2% | 97.2% |
| LMArena Math | 1470 | 1491 |
| FrontierMath (Tiers 1-3) | — | 72.2% |
| FrontierMath Tier 4 | — | 39% |
| ProofBench | — | 87% |
| FrontierMath (Feb 2025 set) | 27.9% | — |
| FrontierMath Tier 4 (v1) | 4.2% | — |
Knowledge Kimi K3 leads
Kimi K2.5: 53.6 (#56), Kimi K3: 63.2 (#21)
| Benchmark | Kimi K2.5 | Kimi K3 |
|---|---|---|
| GPQA Diamond | 87.6% | 93.1% |
| SimpleQA Verified | 34.3% | 50.6% |
| LMArena Expert | 1466 | 1521 |
| Humanity's Last Exam | 24.4% | — |
| Vectara Hallucination Rate | 14.2% | — |
Multimodal Kimi K2.5 leads
Kimi K2.5: 41.1 (#39), Kimi K3: 37.8 (#70)
| Benchmark | Kimi K2.5 | Kimi K3 |
|---|---|---|
| LMArena Vision | 1269 | — |
| Blueprint-Bench 2 | — | 29.5% |
| Furniture Assembly | — | 34.2% |
| LMArena Document | 1430 | — |
Multilingual Kimi K3 leads
Kimi K2.5: 53.9 (#53), Kimi K3: 56.3 (#21)
| Benchmark | Kimi K2.5 | Kimi K3 |
|---|---|---|
| LMArena Non-English | 1433 | 1466 |
| LMArena Chinese | 1495 | 1529 |
| LMArena French | 1454 | 1491 |
| LMArena German | 1441 | 1488 |
| LMArena Japanese | 1421 | 1487 |
| LMArena Korean | 1410 | 1458 |
| LMArena Russian | 1435 | 1482 |
| LMArena Spanish | 1450 | 1472 |
Instruction Following Kimi K3 leads
Kimi K2.5: 75.3 (#64), Kimi K3: 77.7 (#14)
| Benchmark | Kimi K2.5 | Kimi K3 |
|---|---|---|
| LMArena Instruction Following | 1431 | 1483 |
Long Context Kimi K2.5 leads
Kimi K2.5: 52.1 (#7), Kimi K3: 45.8 (#29)
| Benchmark | Kimi K2.5 | Kimi K3 |
|---|---|---|
| LMArena Longer Query | 1445 | 1494 |
| Fiction.LiveBench | 86.1% | — |
| CL-bench | 19.3% | — |
| CL-bench Life | 13.2% | — |
Writing & Preference Kimi K3 leads
Kimi K2.5: 65.1 (#53), Kimi K3: 76.6 (#4)
| Benchmark | Kimi K2.5 | Kimi K3 |
|---|---|---|
| LMArena Text | 1445 | 1476 |
| LMArena Creative Writing | 1423 | 1454 |
| EQ-Bench Creative Writing | 1579 | 2082 |
| LMArena Multi-Turn | 1444 | 1488 |
| EQ-Bench 4 | — | 1339 |
Frequently asked questions
Is Kimi K2.5 better than Kimi K3?
Kimi K3 is the stronger model overall, scoring 59.5 to 48.1 on the Noometry Index. Kimi K2.5 costs 6.7× less per token, which makes it the better buy when Kimi K3's lead doesn't matter for your workload.
Which is cheaper, Kimi K2.5 or Kimi K3?
Kimi K2.5 is cheaper. It lists at $0.45 per million input tokens and $2.25 per million output tokens; Kimi K3 lists at $3 and $15.
Is Kimi K2.5 or Kimi K3 better for coding?
Kimi K3 scores higher on coding benchmarks: 61.0 versus 48.8 in the Noometry coding category.
Which has the bigger context window?
Kimi K3 does, with 1.05M tokens against 262K.
How many benchmarks do Kimi K2.5 and Kimi K3 share?
34 benchmarks have published results for both models. Kimi K2.5 has 51 scored results on Noometry and Kimi K3 has 53.