Fireworks AI Makes Training API Generally Available

Fireworks AI on August 31, 2026 announced the general availability of its Training API and Fireworks Lab, opening its managed training and rollout infrastructure to ML teams that want to run custom training loops on open models. The Training API connects a customer’s own Python training loop to Fireworks-managed distributed compute, covering both the trainer that computes gradients and updates the model and the rollout deployment that generates samples from it.

Under the arrangement described in the announcement, the customer orchestrates the loop from wherever they choose, keeping control over the loss or reward function, the data, and the environment. Fireworks manages the interaction between trainer and rollout, including weight synchronization, failed-swap recovery, and train-rollout alignment. The company positions the API as a response to constraints it says ML teams have reported in existing training workflows: limited model and method choice, restricted control over parameters and training loops, rigid compute, and fragmented training and rollout infrastructure.

Serverless and Dedicated Compute

The Training API offers two compute options. Serverless training lets customers train LoRA adapters on shared infrastructure with per-token billing, with sampling running in the same session; Fireworks describes it as suited to iterating on experiments, de-risking larger runs, or running reinforcement learning loops that alternate between rollout and training. Dedicated training targets full-parameter training, models beyond the serverless pool, larger context lengths or LoRA ranks, and sustained throughput, billed per GPU hour on elastic capacity sized to each run.

The serverless tier attaches to an always-on shared pool with a curated list of popular models, shared capacity, and per-account rate limits. The dedicated tier provisions a trainer and deployment per run, supports LoRA and full-parameter modes up to the largest mixture-of-experts models, and carries no contention or rate limits. Promising checkpoints can be deployed to production inference through the UI or API, and multi-LoRA deployment gives each customer or use case its own tuned model without separate infrastructure, the company said.

Three Training Surfaces

The Training API sits alongside two other surfaces on the Fireworks training platform. Managed Training, aimed at ML engineers, runs built-in jobs in which the customer picks a method — SFT, DPO, or RL — and a base model, launching from the UI or API. Fireworks Lab, also generally available as of the announcement, embeds forward-deployed researchers and engineers with customer teams; engagements open with a diagnosis defining the capability, baseline, success criteria, and scope, then move into a time-boxed implementation, and the customer keeps the production-ready model, evaluation harness, data pipelines, training loop, and recipes.

The Training API itself is aimed at ML researchers and supports SFT, DPO, ORPO, RL, and distillation, from LoRA on serverless compute to full-parameter training on dedicated clusters.

Reinforcement Learning at Scale

Fireworks says it is among the few organizations outside the frontier labs to have operated reinforcement learning across more than 10,000 GPUs, and it frames RL training around three requirements: correctness, performance efficiency, and development speed.

On correctness, the company said the rollout engine and trainer must share the same numerical definition, since small numerical drift can corrupt a learning signal through token clipping or reward collapse while everything appears to be working. Fireworks aligns numeric formats end to end, including BF16, block-wise FP8, and NVFP4, aligns kernels and reduction behavior across both paths, and uses Router Replay to preserve mixture-of-experts routing decisions between rollout and backward pass alongside batch-invariant kernels and deterministic reductions. The company said it validates the alignment by running identical sequences through the rollout engine and trainer and measuring train-inference KL divergence, with continual validation for all models launched on Fireworks training.

On performance, Fireworks said it runs asynchronous RL, overlapping rollout collection with training so rollout GPUs begin generating the next batch while the trainer updates on the previous one, with bounded weight staleness where the algorithm permits. After each training step, updated weights are hot-loaded into the running rollout deployment rather than tearing it down and reloading a full model. For full-parameter checkpoints, the company said it computes an XOR diff between current and previous weights and applies zstd compression, producing up to a 10x reduction in transmission bandwidth.

On development speed, the company described a continuous train, deploy, evaluate, retrain loop on one platform, with checkpoints moving directly into serving and production traces, evals, and feedback feeding the next run. It said teams report two to four times more iterations on the same training budget.

Customer Results Cited

The announcement cited several customers. Harvey post-trained Kimi K3 for long-horizon legal work using asynchronous RL; its Harvey Tenet model scored 19.7% all-pass on LAB against 11.5% for Claude Fable 5, at roughly a third of the cost per task, Fireworks said. Vercel used reinforcement fine-tuning and speculative decoding for v0’s auto-fixer, reaching a 93% error-free generation rate and a 40x end-to-end latency improvement, according to the company. Heidi Health moved its clinical scribe onto fine-tuned open models, going from proof of concept to production in four weeks with 3.5x lower latency. Factory fine-tuned two small LoRA adapters on an open Qwen base to screen for exposed secrets; at a 5% false-alarm budget, the trained model caught about 70% of real secrets versus roughly 59% for GPT-5.5, Fireworks reported.

The Training API is available now through Fireworks’ self-serve signup, with serverless access requiring no provisioning, and the company is directing teams that want hands-on support to Fireworks Lab consultations.