Fireworks AI

Fireworks AI turns the claim that open weights can also be fast into the product: a serverless catalog with published per-token rates (GLM 5.2 at $1.40 per million input and $4.40 output, Kimi K3 at $3/$15, embeddings from $0.008), and below it a deployment ladder — on-demand dedicated GPUs billed per second at $8.00/hr for H100 and H200, $13.00 for B200, $20.00 for GB300, plus reserved capacity with priority access to newer silicon. Training is the other half: LoRA adapters, full-parameter SFT, DPO and reinforcement fine-tuning priced per GPU hour, driven through a Training SDK and the firectl CLI so dataset to endpoint is scriptable, with Nexus as a routing gateway in front of the cheaper paths.

The honest limit is that what Fireworks opens up is open weights and partner models, not the closed frontier; the pitch is latency and unit economics, which is how the company frames itself. Billing splits three ways — per token, per GPU-second, per training hour — so a job that begins as a token count can end as an always-on dedicated deployment, and a dedicated endpoint costs money whether or not traffic arrives. Pick it for fine-tuned models under real traffic, for streaming products where time-to-first-token decides retention, and for teams about to run vLLM or SGLang themselves. Check the autoscaling floor and idle terms on a dedicated deployment, the quantization behind a quoted price, and how catalog deprecations get announced.