Thirteen billion active, and the end of size as a proxy

Same architecture. Better checkpoint. Beats a model with nearly four times the active parameters on every row. Parameter count has been a convenient stand-in for capability for three years, and it just stopped working.

DeepSeek moved V4-Flash out of preview with the same 284B-total, 13B-active architecture it launched with in April — re-post-trained rather than rebuilt — and it now scores above the 49B-active V4-Pro preview on every Terminal-Bench and DSBench-Hard row.

Nothing got bigger

That is the whole finding. No new architecture, no scale-up, no additional experts. The post-training improved and a model activating 13 billion parameters passed one activating 49 billion across the board. Parameter count did not predict the outcome, and parameter count is what most public comparison still runs on.

The practical consequence lands on the bill. Active parameters, not total, set inference cost. A model holding 284 billion in reserve while activating 13 per token is cheap to run and expensive to store — a trade a self-hoster absorbs happily and an API vendor prices around.

The dial is the bigger change

Reasoning effort is now three explicit levels — low, high, max — with a 384K output ceiling at the upper two. Compute per query has become a parameter the caller sets rather than a property the model has.

That quietly breaks benchmark comparison. A score is now a score *at an effort level*, and two labs quoting the same benchmark at different settings are not comparable. Any leaderboard not recording the effort parameter is measuring something it has not defined — and most of them do not record it.

Why the licence is load-bearing

MIT weights on Hugging Face make all of the above checkable. Anyone can download it, run the benchmarks, and disagree in public. That is an entirely different epistemic position from a scorecard published next to an API endpoint, and it is the reason open-weight claims can be argued about productively while closed ones can only be believed or not.

It is also becoming the norm rather than the exception. Three consecutive frontier-class releases have landed on MIT or near-MIT terms.

What to stop doing

Quoting parameter counts as capability. They were always a proxy, they were always imperfect, and this release is a clean demonstration that the proxy has detached from the thing it was standing in for. Ask for active parameters, effort level, and a licence you can check.

Hugging Face — deepseek-ai/DeepSeek-V4-Flash-0731 — model card → · Artificial Analysis — DeepSeek V4 Flash 0731 — intelligence, performance and price analysis → · BenchLM — DeepSeek V4 Flash 0731: benchmarks, MIT weights and pricing →