At Google I/O 2026, Google did something unconventional. Instead of launching a new flagship model centered on intelligence, it launched Gemini 3.5 Flash, a model designed first and foremost for speed.
As models have become capable of taking on more complex coding tasks, the time required to complete a prompt has grown from seconds to minutes, sometimes stretching to hours. As a result, developers are looking for faster inference options. Earlier this year both OpenAI and Anthropic launched high speed variants of their leading models, priced 3x higher than the base model. Google has now joined them, making speed the headline feature rather than an afterthought.
Cerebras is the recognized leader in high speed inference, setting speed records across OpenAI, Kimi, GLM, and Qwen model families. Today we put Google’s fastest model head-to-head against Kimi K2.6 running on Cerebras to see which inference provider can complete tasks the fastest.
Intelligence
Kimi K2.6 is a one-trillion-parameter Mixture-of-Experts model from Moonshot AI, with 32 billion parameters active per token. It is the leading open-weight model among highly capable peers including MiMo V2.5, DeepSeek V4, and GLM-5.1. It is especially popular for coding and is notably used as the base model for Cursor’s Composer 2.5. Gemini 3.5 Flash, by contrast, is a closed model of undisclosed size, designed to run on Google's TPU. Slightly less intelligent than Gemini 3.1 Pro, it is designed first and foremost for speed.
Gemini 3.5 Flash and Kimi K2.6 make for an ideal comparison pair as they both belong in the class of near-frontier models. On the Artificial Analysis Intelligence Index — a composite of ten benchmarks — the two models are neck and neck, scoring 53.9 (Kimi K2.6) and 55.3 (Gemini 3.5 Flash). On coding specifically, Kimi K2.6 stands out. It leads SWE-Bench Pro with a score of 58.6%, ahead of Gemini 3.5 Flash at 55.1%.
The primary measure of inference speed is output tokens per second. The faster the output speed, the faster the model can complete coding tasks. Artificial Analysis tests this with a standard 10,000-token input and measures the rate at which output tokens come back.
Gemini 3.5 Flash achieves 181 tokens/s in this benchmark, significantly faster than Claude Opus 4.8 and GPT-5.5 in the 60 tokens/s range. But Kimi K2.6 on Cerebras is in another category. Cerebras clocks in at 981 output tokens per second — 5.4x faster than Gemini 3.5 Flash. Even against Google's own staged demos, which showed Gemini 3.5 Flash running around 280 tokens per second on what appears to be next-gen TPUs, Cerebras is still more than three times faster. This is achieved by running the model on Cerebras Wafer Scale Engines, which stores the entire model on-chip avoiding the need to load from external memory.
End to End Response
A more comprehensive measure of speed is end-to-end response. It includes input processing, any thinking or reasoning time, and output generation. On the Artificial Analysis measurement (10,000 input tokens, 500 output tokens), Gemini 3.5 Flash completed the task in 17.5 seconds. Kimi K2.6 on Cerebras did it in 5.6 seconds. This shows that even with input processing included, which tends to grow with multi-turn coding tasks, Kimi K2.6 on Cerebras is still able to complete the task in a fraction of the time of 3.5 Flash.
Latency
Voice agents are increasingly being used in customer service, education, and in-car assistants. Latency is by far the most important metric here, with higher latency directly correlating with increased user churn. At 500ms time-to-first-token or more, the conversation starts to feel like a walkie-talkie. The smartest models can take many seconds to respond, which has resulted in developers opting for lesser intelligent models in voice applications.
This tradeoff is no longer necessary. On the latest multi-turn voice agent benchmark (aiewf-eval, by Kwindla), Kimi K2.6 on Cerebras posted the lowest latency in the field at 452ms time-to-first-token — making it the first frontier-class model fast enough for real-time voice. That's a genuine first: a trillion-parameter model clearing the 500ms bar with chain-of-thought reasoning enabled. For comparison, Gemini 3.5 Flash — Google's brand-new, speed-optimized release — comes in at 960ms, and Claude Sonnet 4.6 at 850ms.
Open vs. closed
There's one more dimension that doesn't show up on a benchmark chart. Kimi K2.6 is open. The weights are published under a Modified MIT license, so you can fine-tune it, inspect it, and run it on whatever infrastructure you choose, including Cerebras. Gemini 3.5 Flash is closed and is only available through Google. Even if the model is entirely satisfactory out-of-the-box, there is no second vendor as backup, making it dependent on a single provider’s pricing, deprecation schedule, and uptime.
Conclusion
Every foundational model builder is now offering high speed inference API endpoints. Gemini 3.5 Flash is the fastest of the bunch at 181 tokens/s as measured by Artificial Analysis. Kimi K2.6 on Cerebras matches it on intelligence, generates output five times faster, and completes end-to-end prompts in a third of the time. Moreover, it’s the first frontier model quick enough for real-time voice. Thanks to its open weights, the model can be fine-tuned and deployed as you see fit. Speed and intelligence — you now get both on Cerebras.