Gemma 4 31B vs Kimi K2-Thinking-0905
Gemma 4 31B and Kimi K2-Thinking-0905 are closely matched at 33.1 and 36.0 on the LLM Stats Score. Gemma 4 31B is 5.6x cheaper per token.
Google · Moonshot AI · Updated for 2026
Which is better?
Gemma 4 31B and Kimi K2-Thinking-0905 are closely matched on the overall LLM Stats Score at 33.1 and 36.0.
In the 4 individual benchmarks reported for both models, Kimi K2-Thinking-0905 wins 3; this is a narrower head-to-head signal than the composite indexes.
On price, Gemma 4 31B is roughly 5.6x cheaper per token on a blended 3:1 input/output basis, which adds up quickly at production volume.
Based on current LLM Stats indexes, shared benchmarks, pricing, and model metadata for 2026.
Choose Gemma 4 31B
- cost matters — it's about 5.6x cheaper per token
- you want the most recent training data — it shipped Apr 2026
Choose Kimi K2-Thinking-0905
- you value its reported benchmark strengths — it wins 3 of 4 exact shared results
At a glance
The differences that matter most.
Capability indexes
Additional strengths measured across groups of related public benchmarks
Individual benchmarks
12 reported for Gemma 4 31B · 21 for Kimi K2-Thinking-0905
Gemma 4 31B outperforms in 1 benchmarks (MMLU-Pro), while Kimi K2-Thinking-0905 is better at 3 benchmarks (GPQA, Humanity's Last Exam, LiveCodeBench v6).
Kimi K2-Thinking-0905 shows notably better performance in the majority of benchmarks.
Human preference
Blind head-to-head votes and playground preference scores
Pricing Analysis
Price comparison per million tokens
For input processing, Gemma 4 31B ($0.09/1M tokens) is 5.2x cheaper than Kimi K2-Thinking-0905 ($0.47/1M tokens).
For output processing, Gemma 4 31B ($0.34/1M tokens) is 5.9x cheaper than Kimi K2-Thinking-0905 ($2.00/1M tokens).
In conclusion, Kimi K2-Thinking-0905 is more expensive than Gemma 4 31B.*
* Using a 3:1 ratio of input to output tokens
Model Size
Parameter count comparison
Kimi K2-Thinking-0905 has 969.3B more parameters than Gemma 4 31B, making it 3157.3% larger.
Context Window
Maximum input and output token capacity
Both models have the same input context window of 262,144 tokens. Both models can generate responses up to 262,144 tokens.
Input capabilities
Documented input modalities across available providers
Gemma 4 31B supports multimodal inputs, whereas Kimi K2-Thinking-0905 does not.
Gemma 4 31B can handle both text and other forms of data like images, making it suitable for multimodal applications.
Gemma 4 31B
Kimi K2-Thinking-0905
License
Usage and distribution terms
Gemma 4 31B is licensed under Apache 2.0, while Kimi K2-Thinking-0905 uses MIT.
License differences may affect how you can use these models in commercial or open-source projects.
Apache 2.0
Open weights
MIT
Open weights
Release Timeline
When each model was launched
Gemma 4 31B was released on 2026-04-02, while Kimi K2-Thinking-0905 was released on 2025-09-05.
Gemma 4 31B is 7 months newer than Kimi K2-Thinking-0905.
Apr 2, 2026
5 months ago
6mo newerSep 5, 2025
1.0 years ago
Knowledge Cutoff
When training data ends
Gemma 4 31B has a documented knowledge cutoff of 2025-01-01, while Kimi K2-Thinking-0905's cutoff date is not specified.
We can confirm Gemma 4 31B's training data extends to 2025-01-01, but cannot make a direct comparison without Kimi K2-Thinking-0905's cutoff date.
Jan 2025
—
Provider Availability
Gemma 4 31B is available from DeepInfra, FriendliAI, Novita, Together. Kimi K2-Thinking-0905 is available from DeepInfra, Novita, Fireworks.
Gemma 4 31B
Kimi K2-Thinking-0905
Outputs Comparison
Judge for yourself.
Run your own prompts against Gemma 4 31B and Kimi K2-Thinking-0905 side-by-side, then vote on the output you prefer.
FAQ
Common questions about Gemma 4 31B vs Kimi K2-Thinking-0905.