Fireworks launches Serverless Training API with per-token LoRA pricing
By Rob Litterst, Co-founder, PricingSaaS
Published
Fireworks’ pricing page · Change detected
What changed
Fireworks introduced a Serverless Training API offering per-token LoRA fine-tuning pricing across three models: Qwen 3.5 9B at $0.66 prefill / $1.463 train per 1M tokens, Qwen 3.6 27B at $1.86 / $4.103 per 1M tokens, and Kimi K3 at $10.87 / $32.55 per 1M tokens.
Why it matters
Serverless, per-token training pricing lowers the barrier for teams to customize open-weight models without negotiating enterprise training contracts, which pressures managed fine-tuning vendors to justify their premium with tooling rather than raw compute access. Watch whether Fireworks extends per-token training pricing to closed-weight or larger frontier-scale models, which would signal the metered-training model is becoming the category default rather than a niche offering for smaller open models.
In context
The pricing spread is the real story: training cost scales roughly 2.2x from prefill to train stage within a model, but jumps over 16x from the smallest model (Qwen 3.5 9B) to the largest (Kimi K3), far outpacing the parameter count difference. That signals training compute costs are being passed through with little margin compression at the low end, while Fireworks is likely capturing more markup on flagship-model training where customers have fewer serverless alternatives. This is part of a broader shift among inference platforms — Together, Replicate, and now Fireworks — toward selling fine-tuning as a metered API rather than a managed-service line item.