Skip to content
Vision Index

Kimi K3 vs Qwen3.5-35B-A3B

Kimi K3 leads on the Visual Intelligence Index by 6.2 points, and Kimi K3 scores higher on 6 of the 7 benchmarks where the two differ.

MoonshotKimi K371.9%Rank #5AlibabaQwen3.5-35B-A3B65.7%Rank #7

At a glance

MeasureKimi K3Qwen3.5-35B-A3B
Visual Intelligence Index71.9%65.7%
Rank#5 of 9#7 of 9
Cost per taskMean over the benchmarks, at list price.$0.114—
Latency per taskEach benchmark's median, averaged.29.2s14.3s
Tokens per task31,61925,755
Input / 1M tokens$3.00—
Output / 1M tokens$15.00—
Context window1.05M256K
AvailabilityOpen weightsOpen weights

How cost and latency are measured

Capabilities

Each capability index is a weighted mean of the benchmarks that contribute to the capability.

CapabilityKimi K3Qwen3.5-35B-A3B
Perception71.5%65.3%
Spatial Reasoning73.7%65.8%
Temporal Reasoning72.4%65.5%
Action / Event Understanding66.8%62.2%
Causal Reasoning75.2%70.1%
Information Retrieval70.1%63.6%

Benchmarks

BenchmarkKimi K3Qwen3.5-35B-A3B
Video-MME v255.2%42.5%
LVBench66.8%64.4%
Perception Test82.4%66.7%
NExT-QA86.3%83.4%
Q-Bench Video67.5%66.9%
EgoSchema78.6%77.2%
UCF101-AD39.4%41.3%

Usability

Kimi K3

4.0/ 5

Reachable1.0

There is an endpoint you can call without hosting anything.

  • First-party API — supported
  • Third-party API — supported
Portable1.0

You can run it yourself, and are not tied to one vendor.

  • Open weights — supported
  • Self-hostable — supported
Multimodal input0.5

It takes the footage directly, rather than frames you extracted.

  • Video ingestion — supported
  • Audio — not supported
Programmable1.0

Output you can parse, and tools it can call on its own.

  • Structured output — supported
  • Tool calling — supported
Operable0.5

Usable interactively and in bulk, not only one call at a time.

  • Streaming — supported
  • Batch API — not supported

Qwen3.5-35B-A3B

4.5/ 5

Reachable1.0

There is an endpoint you can call without hosting anything.

  • First-party API — supported
  • Third-party API — supported
Portable1.0

You can run it yourself, and are not tied to one vendor.

  • Open weights — supported
  • Self-hostable — supported
Multimodal input1.0

It takes the footage directly, rather than frames you extracted.

  • Video ingestion — supported
  • Audio — supported
Programmable1.0

Output you can parse, and tools it can call on its own.

  • Structured output — supported
  • Tool calling — supported
Operable0.5

Usable interactively and in bulk, not only one call at a time.

  • Streaming — supported
  • Batch API — not supported