Ember-1 vs. Kimi K3: Nearly identical results at 3.4 times the speed

The recent launch of Ember-1, a research preview of Fireworks built on Moonshot's open-weight model, has demonstrated nearly identical results to Kimi K3, but with nearly 40% fewer tokens. According to Fireworks, Ember-1 learns to cut unnecessary reasoning while retaining crucial thinking, resulting in cost savings. However, the pricing for both models on OpenRouter varies significantly, with Ember-1 costing $3 per million input tokens and $15 per million output tokens, while Kimi K3 can be as low as $1 per million input tokens and $9 per million output tokens. Fireworks charged the same for both models, so the comparison is fair.

To determine if Ember-1's reduced reasoning truly impacts accuracy, Fireworks tested both models on a series of progressively difficult tasks, including logic puzzles, deploy scheduling, and probability questions. Both models consistently solved the logic puzzles and deployed the correct schedule, with Ember-1 averaging significantly fewer reasoning tokens and completing tasks much faster.

In the probability test, Ember-1 made a minor arithmetic error on one run out of 15, but overall, it performed almost identically to Kimi K3.

On average, Ember-1 completed the tests 3.4 times faster than Kimi K3 and used 23% fewer reasoning tokens. This resulted in a total cost of $2.48 for Ember-1 compared to Kimi K3's $3.26 on Fireworks, making Ember-1 24% cheaper. However, Kimi K3's actual cost can be lower depending on the provider, with a minimum price of $1.96.

In conclusion, Ember-1 is nearly as accurate as Kimi K3 while offering improved speed and reduced costs. If the goal is to save money while maintaining high accuracy, Ember-1 is the clear choice. On the other hand, if cost is not a primary concern and speed is less important, Kimi K3 may still be a viable option, especially when using a cheaper provider on OpenRouter.