Kimi K3 is one of the best open models you can specialize. But running RL on it is where most teams stall. The hard part is everything in the loop: training steps, rollouts, moving weights between them, and keeping both sides numerically aligned. If any one of those is slow or wrong, the whole run is slow or wrong. Our research team just published how we make RL on K3 practical on @FireworksAI_HQ: K3 mixes linear attention, latent attention, and an MoE with 896 experts. Generic kernels leave GPUs idle while experts wait on communication. Our MoE kernel overlaps that communication with compute, delivering 1.87–2.61x faster forward passes on the expert layers than an already-optimized DeepEP baseline. At 128K tokens, our attention-residual kernels cut peak memory by 44%. Rollouts are where RL time goes. RL rollouts aren’t chat. Thousands of trajectories branch from shared prompts, call tools, and grow for 50 turns. K3 also keeps two kinds of cached state, which have to stay in sync whenever a trajectory branches. We coordinate both and hit a ~99% cache hit rate on long multi-turn rollouts. Speculative decoding can corrupt your state. On a recurrent model, a rejected draft token can leave the state wrong. We replay only accepted tokens from a committed checkpoint. In separate tests, that alone raised rollout throughput by 65%. Also, 1TB checkpoint can’t move across the cluster every update. We send compressed XOR deltas instead, which are 98% smaller per update and rebuild the new weights bit for bit. Finally, we enforce deterministic kernels, use batch-invariant normalization, and replay the rollout’s expert choices during training. With all three, train–rollout KL divergence was as low as 0.0008 in our synchronous LoRA setup. That’s the difference between a run that learns and one that only looks like it’s learning. None of this should be your team’s job. With serverless training, your experiment plugs into an already-running pool of K3 trainers and rollout workers. You don’t provision GPUs, launch jobs, or bring up deployments. You bring tasks, an environment, and a reward function in Python. The teams that win with RL aren’t the ones with the biggest clusters. They’re the ones that iterate fastest on reward and data. We have supported LoRA and Full Param training for a while now and can't wait to add Kimi 3.1. fireworks.ai/research/kimi-k…

@tryheidi’s 12-person team post-trained @nvidia’s open Nemotron 3 Ultra, and clinicians prefer it to the frontier. Their blog is one of the clearest cases I’ve seen for owning your model in a regulated industry. Heidi supports 2.7M clinical visits a week. Their Evidence product ran on a frontier model and worked well. But on a closed model, the only fix for a weakness clinicians flagged was a prompt edit. And healthcare needs pinned versions and on-prem deployment. So they post-trained Nemotron on @FireworksAI_HQ and tested it in blind side-by-side comparisons with clinicians against the production frontier model. What I’d take from it: Pick the model for your workload. Requests carry tens of thousands of tokens. Nemotron has only a few attention layers, so a long request needs a fraction of the memory a similar transformer would. And it was pretrained in NVFP4. Never let training see what production won’t. All training data was synthetic. The generator could see reference answers; the training prompt couldn’t. Gates before grades. Any safety or faithfulness failure gets the minimum reward. Otherwise the policy trades faithfulness for completeness. Safety offline, preference online. A clinician can’t catch an unsafe answer in the response they didn’t pick. Every checkpoint passes a safety eval first. My favorite detail: the 4-bit model matched BF16 on 97.35% of tokens on real production conversations. Most teams underestimate this. Numerics differ between stacks, and in RL it compounds. If your trainer and rollout engine disagree on token probabilities, the policy learns from samples it didn’t actually produce. That’s why we run training and inference on one stack and keep the KL divergence between them low. Heidi trained, ran rollouts, and deploys on the same platform. Heidi’s team owned data, reward, and evals. Fireworks provided the SOTA infra with dedicated Blackwell deployment. That’s the split you want. Your people on the problem only you can solve, us on the infrastructure that eats months when you build it yourself. Renting a model gets you a prompt box. Owning one gets you a model you can actually fix when your users tell you what’s wrong. How Heidi did it → fireworks.ai/customers/heidi…

have said it before and will say it again, not all tokens are created equal. love how the team does not compromise on quality.