The Pulse
Fireworks Brings Ember-1 to Vercel With 40% Fewer Tokens
Fireworks’ Ember-1 model, built on Kimi K3, used 39% fewer total tokens in one customer coding test while scoring slightly higher, the company says. Vercel added the model to AI Gateway on September 27, with a two-week research preview.
AI.info Team · Written with AI · fact-checked ·
In one customer’s production coding test, Fireworks says Ember-1 used 39% fewer total tokens than Kimi K3 while scoring slightly higher. The new model became available through Vercel’s AI Gateway on September 27, giving developers another option for coding agents that make repeated model calls.
Fireworks built Ember-1 on Kimi K3 and says it trims unnecessary reasoning without sacrificing comparable quality. The company is pitching lower token use as a way to reduce costs in agent workflows, where earlier reasoning can be included again in later calls.
One customer test cut total tokens by 39%
In the customer comparison Fireworks published, Kimi K3 scored 0.751, took an average of 23.8 steps and generated 49,300 output tokens per task. Ember-1 scored 0.753, took 21.4 steps and generated 29,900 output tokens. For that test, reasoning tokens were reduced by 71.3% and total tokens by 39%, Fireworks reports.
Two customers tested Ember-1 on production coding workloads, with about 35% fewer tokens per task at comparable quality, the company says. One of those customers has moved the model into live production and plans to expand its use, according to Fireworks. Those results come from the company’s tests; they do not establish that every coding agent or task will see the same savings.
Shorter reasoning is the product bet
Fireworks ran more than 50 training experiments and over 200 evaluations to teach the model to produce shorter reasoning traces. Its argument is that long traces carry extra expense in multi-turn agentic work: each new turn can bring earlier context back into the prompt, increasing the tokens processed across a task.
“The most cost optimized way to run K3 is no longer to make it think less, but to run Ember-1, the model that learned to think efficiently.” Fireworks
Fireworks also compared Ember-1 with Kimi K3 across coding benchmarks. On SWE-bench Verified, its published results show 92.2% for Ember-1 versus 93.2% for K3 at maximum reasoning effort; on Terminal Bench 2.1, Ember-1 scored 82.0% versus K3’s 80.9%. The company’s table reports lower benchmark costs for Ember-1, but the figures are Fireworks’ evaluation results, not an independent audit.
Vercel’s gateway adds a two-week preview
Vercel’s September 27 changelog lists Ember-1 as a research preview with an initial two-week availability window. The model supports a 1-million-token context window, text and image input, and tool calling. Developers can select fireworks/ember-1 in compatible coding agents after setting up AI Gateway.
Fireworks lists serverless rates of $3 per million input tokens and $15 per million output tokens. The cost case, then, rests on generating fewer tokens rather than a lower posted rate: whether that translates into a lower bill depends on an agent’s workload and whether it completes the same work with similar quality.