Train past the frontier: Training API now generally available

The next frontier of competitive advantage is specialized intelligence. The most ambitious companies are training models with the capabilities that underpin differentiated products and operations, across coding, chip design, cybersecurity, and beyond:

• Harvey post-trained Kimi K3 for long-horizon legal work using asynchronous RL. Harvey Tenet scored 19.7% all-pass on LAB against 10.8% for base Kimi K3 at effectively flat cost per task.
• Vercel used RFT and speculative decoding to fine-tune an open model for v0’s auto-fixer, reaching a 93% error-free generation rate for v0 and 40x end-to-end latency improvement.
• Heidi moved its clinical scribe onto open models it fine-tuned, going from POC to production in four weeks and achieving 3.5x lower latency.
• Factory fine-tuned two small LoRA adapters (one to flag risky secrets, one to clear false alarms) on an open Qwen base. At a 5% false-alarm budget, the trained model caught about 70% of real secrets versus ~59% for GPT-5.5, while being faster and more cost-effective.

Across these workloads, training a base model for specific tasks improved quality, latency, cost, and/or behavioral consistency. Fireworks makes this level of specialization accessible to teams through three training surfaces, each designed for a different level of expertise, operational ownership, and control.

Open in lightbox

Today, we are announcing the general availability of our Training API and Fireworks Lab.

We’ve worked with ML teams around the world to understand where existing training workflows fall short: constrained model and method choice, limited control over parameters and training loops, rigid and underutilized compute, and fragmented training and rollout infrastructure that bloats costs, creates synchronization overhead, and introduces numerical drift. Those partnerships shaped our Training API.

The Fireworks Training API connects your custom training loop to Fireworks-managed distributed training and rollout infrastructure. You orchestrate the loop in Python from wherever you choose, with full control over your loss or reward, data, and environment. Fireworks manages high-performance compute: the trainer computing gradients and updating the model, and the rollout deployment generating samples from the trained model. Fireworks also manages the interaction between these two, from weight synchronization to failed-swap recovery, and train-rollout alignment. All of this allows you to focus on refining your learning signal, not infrastructure challenges.

1234567
def reward_fn(completion: str, row: dict) -> float:
    """Return 1.0 if the model's numeric answer matches the ground truth."""
    predicted = extract_answer(completion)
    truth = extract_answer(str(row.get("ground_truth", "")))
    if predicted is None or truth is None:
        return 0.0
    return 1.0 if predicted == truth else 0.0

Example of a custom reward function defined in Python. This code can be dropped directly into a training loop using the Fireworks Training API.

Choose Serverless when speed to experiment matters most, or talk to us about Dedicated when scale, control, and GPU economics take priority.

Serverless training allows you to train LoRA adapters on top of shared infrastructure. You pay per token, and sampling runs in the same session. Use it to iterate on experiments, de-risk a larger run, or run RL loops that alternate between rollout and training.

Start on Serverless in minutes, nothing to spin up: Start training →

Open in lightbox

Dedicated training* is best when you need full-parameter training, models beyond the serverless pool, larger context lengths or LoRA ranks, or sustained throughput where per-GPU-hour pricing is more economical than per-token. Fireworks gives you dedicated, elastic capacity sized to each run, with the ability to scale GPU type and cluster size as needed.

*Requires Fireworks Lab consultation and provisioning

Interested in Dedicated? Talk to us →

Fireworks is among the few organizations outside the frontier labs to have operated reinforcement learning (RL) across more than 10,000 GPUs. RL turns training and inference into one continuously coupled system; every step starts with rollouts from the current policy, updates the weights, and feeds those weights back into inference for the next round of sampling. At frontier scale, the algorithm is only part of the challenge. Effective RL hinges on three requirements: correctness, performance efficiency, and development speed.

For correctness in RL training, the rollout engine and trainer need to share the same numerical definition. Small numerical drift can corrupt a learning signal, from token clipping to reward collapse, when everything appears to be working. With Fireworks,

• Numerical formats match end to end. Differences in precision or quantization between training and inference can change the values the model computes, even with the same weights. Fireworks aligns trainer and rollout numeric formats (e.g. BF16, block-wise FP8, and NVFP4) so the model the trainer saves is the one that the rollout deployment runs.
• Kernels match, too. The same numeric format does not guarantee the same computation. Floating-point reductions, kernel selection, and parallelism can produce different numbers even at identical precision, so Fireworks aligns the underlying implementations across both paths.
• MoE execution stays consistent. Small numerical differences can change which experts fire and send a token down a different computational path. Fireworks uses Router Replay to preserve expert selections between rollout and backward pass, alongside batch-invariant kernels and deterministic reductions for supported frontier MoEs.

Fireworks validates this alignment directly by running the same sequences through the rollout engine and trainer and measuring train-inference KL divergence (KLD). All models launched on Fireworks training are continually validated to ensure KLD is minimized.

Open in lightbox

At scale, RL is largely a generation problem. Much of the wall-clock time is spent producing rollouts rather than updating weights, so rollout throughput and GPU utilization become primary determinants of training time and cost. Completion lengths also vary widely, which means rigid batching can leave GPUs idle while a small number of long-running trajectories finish.

Fireworks can run asynchronous RL, overlapping rollout collection with training rather than forcing the two sides of the loop to execute sequentially. Rollout GPUs can begin generating the next batch while the trainer updates on the previous one, with bounded weight staleness available when the algorithm permits more overlap. The objective is simple: keep GPUs doing useful work instead of waiting at synchronization barriers.

Once the loop completes a training step, the updated model must be transmitted and loaded into the rollout deployment. Rather than tearing down the rollout deployment and reloading a full model after every step, Fireworks hot-loads new weights into the running deployment. For full parameter checkpoints, novel compression techniques are used to reduce the bandwidth and storage requirements for weight transmission. An XOR Diff is computed between current and previous weights, and zstd compression is used to reduce the size of this payload. This results in up to a 10x reduction in transmission bandwidth, significantly speeding up the training loop.

The bottleneck to production is often not the training job, but the time between iterations. In a fragmented stack, every cycle means moving checkpoints between training and inference systems, redeploying, reconciling differences, running evals, and starting again.

Fireworks turns that process into a continuous development loop: train → deploy → evaluate → retrain on one platform. Checkpoints move directly into serving, while production traces, evals, and feedback feed the next training run. And because we handle GPU provisioning, capacity, and configuration, customers skip the procurement, model enablement, and infrastructure-sizing work that slows iteration and time to market. Teams report 2–4x more iterations on the same training budget, compressing model development from weeks to hours and accelerating time to production.

Open in lightbox

With correctness, performance efficiency, and iteration speed built into the platform, teams can focus on their training strategy. The strategy includes the decisions that shape model quality: the training loop, data, rewards, evaluators, and harness design. It also includes the systems choices that determine how the workload should run at scale, such as synchronous versus asynchronous RL, acceptable weight staleness, rollout-to-trainer allocation, parallelism, and checkpoint and hot-load strategy.

Fireworks Lab helps teams make those decisions more confidently and execute faster. Forward-deployed researchers and engineers embed directly with your team. A full engagement starts with a diagnosis that defines the capability, baseline, success criteria, and scope, then moves into a time-boxed implementation built backward from your target date. You keep the production-ready model, evaluation harness, data pipelines, training loop, and recipes.

Talk to Fireworks Lab →