Baseten pushes DeepSeek V4.1 Flash to 991 tokens/s
Baseten says it used an internal agent harness to optimize more than 80 kernels for DeepSeek V4.1 Flash 1017, producing a "Fast" variant that Artificial Analysis lists at 991 output tokens per second.
Baseten detailed how it built a faster serving variant of DeepSeek V4.1 Flash 1017. The company says it used an internal agent harness to optimize more than 80 kernels and bottlenecks. According to Baseten, the resulting Fast variant appears on Artificial Analysis at 991 output tokens/s, with the lowest time per task on the frontier.
The base model is a 552B-parameter mixture-of-experts at native precision. It activates 8B parameters during prefill and 16B during decode. Baseten says it scores 32 on the Artificial Analysis V5.0 Index, matching GLM-5.3 at roughly one-third the cost and four times the speed. The post also cites Cline's testing, in which the model outperformed 5.6 Sol on Terminal-Bench 4.0 at about 17x lower cost.
Baseten argues that kernel work suits agents because profiling, rewriting and benchmarking are well-defined and verifiable. It named three optimizations close to the hardware layer:
- Kernel fusion, which keeps intermediate results in on-chip memory instead of passing them through VRAM. Examples include a fused QK-norm/RoPE kernel and an improved fused MoE router kernel that Baseten says cuts decode time by 2.7ms.
- mHC (Manifold-Constrained Hyper-Connection), the residual-stream design DeepSeek's model uses. Baseten says it improved the open-source mHC megakernel in DeepInfer, making it 30% faster at batch size 1.
- Programmatic Dependent Launch and host loop minimization. The excerpt available cuts off before these sections, so their specific gains aren't known.
Baseten says it plans to release open-source work on each technique soon. It has not given a timeline.