Fireworks offers zero-KLD RL training infrastructure as managed service for GLM 5.2

Fireworks AI now offers a managed service for reinforcement learning on GLM 5.2 that keeps training and serving numerically identical, with zero KLD end to end. The service is available through the Fireworks Training API, with the numerical foundation carried forward from GLM 5.1. The validated training shape is public.

The hard part of RL on a frontier model is not the algorithm but the infrastructure that makes training and inference produce the same numbers. Fireworks says this infrastructure, previously confined to top frontier labs, is now available as a managed service. The service includes batch invariance for large mixture-of-experts models, meaning a request returns the same result regardless of other traffic in its batch. Without it, an on-policy run becomes off-policy because the rollout was generated by a subtly different model than the one being updated.

Zero-KLD train/serve alignment

For GLM 5.2, the usual fixes are insufficient. The popular Mixture-of-Experts fix, router replay, handles which experts fire but cannot address which tokens the sparse-attention indexer selects. Those selections are not tractable to replay. So the generation engine, the prompt-reading path, and the trainer must share one numerical definition. Fireworks says this is the first industry offering of whole-system end-to-end invariance for a frontier MoE like GLM 5.2.

The challenge stems from floating-point math where addition is not associative. A frontier MoE changes the order of additions constantly due to latent attention, the sparse indexer, expert matmul kernels, router near-ties, and cross-GPU all-reduce algorithms. Stacked together, temperature 0 on a busy server is quietly nondeterministic. Fireworks pinned every one of those decisions so they depend only on the request, without giving up so much speed that the trainer becomes unusable for everyday SFT/DPO.

Validation and performance

Fireworks ran the GLM countdown reasoning task two ways. Without the Fireworks numerics stack, train-inference KL was around 0.013 and drifting, about 45% of tokens were clipped, and reward collapsed around step 20 from around 0.9 to under 0.2. With the stack, train-inference KL was 0, zero tokens were clipped, and reward stayed healthy across the full 25-step run. Fireworks calls importance sampling and clipping a tax, not a fix, because every clipped token is learning signal thrown away.

Open-source deterministic modes typically run 35-60% slower, but Fireworks says it pays virtually none of that tax. The GLM trainer holds around 3,500 tokens per second per node, on par with the OSS TileLang implementation, and layers zero-KLD numerics on top. Rollouts generate about 1.8x faster on GLM 5.2 than GLM 5.1, around 5,000 tokens per second per node, per promotion-CI.

What the service includes

Fireworks offers SFT, DPO, and RL through the Training API, with SFT and DPO on managed training. On-policy RL is available where the trainer and serving engine genuinely agree. Temperature-0 requests return the same answer regardless of server load, enabling reproducible evals and regression tests. The full loop can run on managed infrastructure with trainer and deployment co-located for fast weight sync, or longer runs can be driven through the API.

The service targets teams that want to post-train and reinforce frontier models on the same infrastructure top labs built in-house. Fireworks frames it as frontier specialized intelligence as a service, not a frontier you rent prompts from. The company also announced its inaugural conference, Forge 2026.