News · 2026-08-12

DeepSeek's new open model is 1.6 trillion parameters and runs 49 billion of them per token

DeepSeek published DeepSeek-V4-Pro to Hugging Face with 1.6 trillion total parameters, 49 billion activated per token, and a one-million-token context window. It is the only openly downloadable model among this week's three frontier-tier releases, sitting alongside Alibaba's hosted-only Qwen3.8-Max and xAI's hosted-only Grok 4.6. The gap between the two numbers on the model card is the whole engineering story.

Key facts

• 1.6 trillion total parameters with 49 billion activated per token, per the DeepSeek-V4-Pro model card.
• A one-million-token context window, published under an MIT license.
• Exposed as deepseek-v4-pro in DeepSeek's API changelog and pricing docs.
• The Hacker News discussion turned within an hour from capability to whether anyone outside a datacenter can run it.

What the two numbers mean

DeepSeek's model card lists the model at "1.6T total parameters, 49B activated parameters." Those describe entirely different costs, and conflating them is the single most common mistake people make reading a model release.

The 1.6 trillion is how much model exists. The 49 billion is how much of it runs when you feed in a token. This is a mixture of experts: the network is carved into many specialist sub-networks, and a small router decides which handful to wake up for each token. Everything else stays asleep.

The analogy that holds up is a hospital. A large teaching hospital employs thousands of specialists, and its payroll scales with all of them. But a patient walking through the door sees a triage nurse and two doctors. The cost of treating that patient scales with the three people in the room, not the thousand on staff. What the hospital's size buys is that whichever specialist you happen to need is already employed.

So compute per token scales with 49 billion, which is why DeepSeek can serve this at a price that is not absurd. Memory and storage scale with all 1.6 trillion, which is why you cannot run it on your desk.

The part that runs into physics

Commenters on Hacker News reached the practical objection almost immediately, and they are right. Even quantized, a 1.6-trillion-parameter weight store lands in the hundreds of gigabytes. It has to be resident somewhere the model can reach fast enough, because LLM inference is bounded by memory bandwidth long before it is bounded by arithmetic. A sparse model does not escape that; it just changes which numbers you have to move.

There are partial routes around it. Aggressive quantization shrinks the store at some cost to quality, and offloading and streaming weights trades speed for capacity by keeping most of the model on slower storage. Neither turns this into a laptop model. The realistic profile is a multi-GPU server or a rented cluster, which puts genuine local deployment in the hands of labs, well-funded startups, and hobbyists with unusual hardware budgets.

Why "open" still matters here

It would be easy to dismiss an open release nobody can run as a gesture. It is not, for three reasons.

The weights are auditable. Anyone can inspect what is actually in the model rather than trusting a system card about it. That matters for interpretability research, for safety evaluation done by people who do not work at the lab, and for anyone trying to establish what a model was trained on.

The weights do not expire. A hosted model can be deprecated, rate-limited, repriced, or silently swapped for a newer version that behaves differently. A checkpoint on disk cannot. For regulated deployments where the exact model version is part of a compliance record, that distinction is the entire ballgame. Our note on what open weights really means works through where the line actually falls.

And the weights get distilled, quantized, and pruned by the community into things that do fit on smaller hardware. That is the pattern the open ecosystem has followed every time, most recently after Ant put Ling 3.0 Flash online under plain MIT and after two-bit models started running on every major llama.cpp backend. The 1.6-trillion checkpoint is the source material, not the product.

The caveat

DeepSeek has not published a detailed technical report alongside the weights at the level of its earlier releases, and the practical performance claims circulating around this model are mostly community benchmarks rather than a lab-issued system card. Reported numbers on math and proof tasks look strong in the Hacker News thread, but the same thread is candid that benchmark wins have repeatedly failed to survive contact with real workloads. Treat the parameter counts and the license as verified, and the capability claims as provisional until independent evaluations land.

Primary source, verified: read the paper →

Ground Truth. (2026, August 12). DeepSeek's new open model is 1.6 trillion parameters and runs 49 billion of them per token. Ground Truth. https://groundtruth.day/news/deepseek-put-a-1-6-trillion-parameter-model-on-hugging-face.html

@misc{groundtruth:deepseek-put-a-1-6-trillion-parameter-model-on-hugging-face,
  title  = {DeepSeek's new open model is 1.6 trillion parameters and runs 49 billion of them per token},
  author = {{Ground Truth}},
  year   = {2026},
  month  = {aug},
  url    = {https://groundtruth.day/news/deepseek-put-a-1-6-trillion-parameter-model-on-hugging-face.html}
}

Topics: open-weights · mixture-of-experts · deepseek · model-release · local-ai · inference