Correction, . This note originally said 1.7T, following an early Baseten write-up. The model card, NVIDIA's NIM card and the V4 paper all say 1.6T total, 49B active. Corrected throughout. The frontier page carries the sourced numbers.
DeepSeek V4 Pro 0813 is a 1.6T-parameter open frontier model under the MIT license. That's the sentence that changes the serving conversation: a model this size, open, and permissively licensed means the frontier is now something you can run yourself.
I keep this note mostly as a record of how fast the ground moved. Writing it, the interesting fact was the parameter count. In hindsight the interesting fact was the licence, and I did not weight that nearly heavily enough at the time.
The model
1.6T parameters, MIT license, open weights. DeepSeek's MoE architecture means the 1.6T parameters are active sparsely, which is what makes serving it feasible. The model is available for inference today on Baseten's model APIs, and it's the kind of model that makes the "open frontier" phrase real.
What it takes to serve
Serving a 1.6T-parameter MoE is a different problem from serving a 70B dense model:
- Disaggregation. Prefill and decode need different hardware, because the memory and compute profiles diverge.
- Expert parallelism. The experts have to be spread across GPUs, and routing decides which GPUs light up.
- Quantization. FP8/FP4 is not optional at this scale; it's the difference between feasible and not.
- KV cache. Long context means the KV cache dwarfs the weights, and cache management is the new bottleneck.
Serving a 1.6T MoE is like running a city where only a few neighborhoods are awake at any given time. The routing decides which neighborhoods, and the infrastructure has to serve all of them.
What an open 1.6T actually changes
- Open frontier. A 1.6T MIT-licensed model means the frontier is reproducible, not just consumable.
- Cost. Frontier intelligence at a fraction of the cost of closed APIs, if you can serve it.
- Inference engineering. This is the workload that makes the whole field matter: the model is open, so the differentiator is how well you serve it.
When a 1.6T model is open and MIT-licensed, the moat isn't the model anymore. It's the serving.
Sources
- Baseten: Inference engineering for DeepSeek V4 Pro 0813: the serving story and the 1.6T number.
- Together AI: DeepSeek V4 Pro 0813 vs GPT-5.6-SOL: the cost, coding, and routing comparison.
- DeepSeek docs: the model card and API.
- DeepSpec repo: the speculative decoding stack that makes it fast.