Prime Intellect has released Prime Mega MoE, a training extension of DeepSeek’s fused Mega MoE kernel that adds a fast, overlapping backward pass. This makes the fused kernel usable for training, not just inference. In its prime-rl trainer, full training steps on Qwen3 MoE models run up to 34% faster. Against TorchTitan, the gain reaches 55%. For teams training mixture-of-experts models, that is a direct cut in GPU hours.
TL;DR
- Size: A GPU kernel, not a model. Tested on Qwen3-30B-A3B and Qwen3-235B-A22B (128 experts, top-8 routing).
- Runs on: NVIDIA Hopper (blockwise FP8) and Blackwell (MXFP8) GPUs, BF16 on both. Published numbers use 8×H200 nodes, 32 to 128 GPUs.
- Performance: 32% to 34% faster full training steps than the prime-rl baseline. 48% to 55% faster than TorchTitan.
- Best: 3.11x faster forward pass on a Qwen3-30B-A3B MoE layer (12.0 ms to 3.9 ms).
- Worst: 1.24x backward pass on a Qwen3-235B-A22B MoE layer (30.5 ms to 24.6 ms).
What is Prime Mega MoE?
Prime Mega MoE is a fused, overlapping MoE kernel for training from Prime Intellect. It builds on the Mega MoE kernel in DeepSeek’s open-source DeepGEMM library.
Why is MoE training communication-bound?
In an MoE model, each token goes to a small subset of experts. Those experts sit on different GPUs. So every layer sends tokens out, runs expert math, and sends results back. This happens in the forward pass and again in the backward pass.
In a standard setup, these are separate operations. The GPU sorts tokens and waits on the transport layer. Then it computes, hands results back, and waits again. Those waits are idle GPU time.
How does the fused kernel work?
DeepSeek’s Mega MoE, added to DeepGEMM in April 2026, replaces that chain with one kernel call. It fuses expert-parallel dispatch, the 2 expert linear layers, SwiGLU, and combine. While one token group is in transit, the kernel computes on a group that already arrived. It also skips intermediate tensors by using GPU-only scratch buffers. But it only shipped a forward implementation.
Prime Intellect wrote the missing backward kernel, also overlapping. Supported types are BF16, MXFP8 (Blackwell only) and blockwise FP8 (Hopper only). The prime-rl config also exposes an fp8_bwd mode: BF16 forward, FP8 backward.
How fast is Prime Mega MoE?
All numbers below are self-reported by Prime Intellect. Training steps use 8×H200 nodes, FP8, 16,384-token sequences and full activation checkpointing. Each result is the mean of steps 6 to 20.
Full training step, seconds (lower is better):
| Model (cluster) | prime-rl baseline | TorchTitan (FP8) | Prime Mega MoE | Gain vs baseline | Gain vs TorchTitan |
|---|---|---|---|---|---|
| Qwen3-30B-A3B (32 H200) | 2.78 | 3.07 | 2.08 | 34% | 48% |
| Qwen3-235B-A22B (128 H200) | 12.72 | 14.98 | 9.65 | 32% | 55% |
Against TorchTitan, throughput rose from 5,337 to 7,886 tokens/s/GPU on the 30B model. On the 235B model, it rose from 1,094 to 1,697. The prime-rl baseline uses torch all-to-all plus DeepGEMM FP8 experts.
Single MoE layer, milliseconds (1 node, 8×H200, EP=8, 128 experts, top-8, random routing):
| Layer | Forward | Backward | Full layer (2x fwd + bwd) |
|---|---|---|---|
| Qwen3-30B-A3B | 12.0 to 3.9 (3.11x) | 15.4 to 8.2 (1.87x) | 39.3 to 15.9 (2.47x) |
| Qwen3-235B-A22B | 20.7 to 9.5 (2.17x) | 30.5 to 24.6 (1.24x) | 71.9 to 43.7 (1.64x) |
The pattern is clear. Forward gains are large, and backward gains are smaller. The headline “up to 3.11x” refers to the forward pass only.
How does it compare with other fused MoE options?
| Feature | Prime Mega MoE | DeepGEMM Mega MoE | TorchTitan + DeepEP | AMD Primus MegaMoE |
|---|---|---|---|---|
| Developer | Prime Intellect | DeepSeek | PyTorch, Nebius | AMD |
| Training (backward) | Yes, fused and overlapping | No, forward only | Yes (separate dispatch kernels) | Yes, per design docs |
| Compute and comms fused | Yes | Yes | No | Yes (dispatch+GEMM1, GEMM2+combine) |
| GPUs | Hopper, Blackwell | SM90, SM100 (library) | B200 in published runs | MI355X (gfx950) |
| Data types | BF16, MXFP8, blockwise FP8 | FP8xFP8, FP8xFP4 | BF16, MXFP8 grouped GEMMs | BF16, MXFP8 |
| Published speedup | Up to 55% step time vs TorchTitan | Not disclosed | Up to 41% on DeepSeek-V3 671B | Not disclosed |
| License | Apache 2.0 (prime-rl) | MIT | BSD-3-Clause | MIT |
TorchTitan’s 41% gain comes from a different setup (B200, 671B model, BF16 baseline). It is not directly comparable.
How do developers use it?
Prime Mega MoE is a dispatch backend in prime-rl. Set type = “mega_moe” under [trainer.model.moe.dispatch]. Choose precision from bf16, mxfp8, fp8 or fp8_bwd. The training kernels ship in a prebuilt DeepGEMM wheel from prime-kernels. The stock DeepGEMM build lacks them.
What it means
Engineers training large models will see faster iteration cycles, particularly when using Hopper or Blackwell hardware. The backward pass improvement is modest compared to the forward pass, but the combined step time reduction is significant enough to justify switching backends for production training runs.
