Iro AIIro AI Blog
DeepSeek V4 is the strongest open-weight model of 2026: 1.6 trillion parameters, a million-token context, MIT licensed, free to download. Almost none of that matters if you are not running your own servers — so here is the part that does.
DeepSeek V4 is the most capable openly licensed model released in 2026, and for most people its importance is economic rather than practical. You are unlikely to run it. You will probably still benefit from it existing.
The verifiable facts, from the model cards themselves: two sizes, V4-Pro at 1.6T total parameters with 49B active and V4-Flash at 284B total with 13B active, both with a 1M-token context window, both released under the MIT licence with weights on Hugging Face. V4-Flash-0731 was promoted to a production-candidate build on 31 July 2026.
A note on numbers you will see elsewhere. Coverage of this release has been unusually sloppy — parameter counts off by a factor of ten, and API availability reported as generally available while DeepSeek's own docs still listed it as preview. Everything above comes from the model cards and DeepSeek's API documentation. Where third-party write-ups disagree with those, trust the source.
| Total parameters | Active per token | Context | Licence | |
|---|---|---|---|---|
| V4-Pro | 1.6T | 49B | 1M tokens | MIT |
| V4-Flash | 284B | 13B | 1M tokens | MIT |
The two-number format trips people up, so it is worth unpacking. These are mixture-of-experts models: the full network is enormous, but only a fraction of it fires for any given token. V4-Pro holds 1.6 trillion parameters and uses about 49 billion of them at a time.
That design buys speed and cost, not a smaller footprint. The whole model still has to be resident in memory, because the router cannot know in advance which experts a token will need. Active parameters tell you what it costs to run. Total parameters tell you what it costs to host. Marketing tends to quote whichever is more flattering.
Iro AI turns ideas like the ones in this post into 5-minute exercises with feedback. Free tier, Pro from $4.17 a month ($49.99 a year, 7-day free trial).
"Open" gets used very loosely in AI. Plenty of models described as open ship with licences restricting commercial use, capping your user count, forbidding using outputs to train other models, or reserving the right to change terms later.
MIT has essentially none of that. Download it, run it, modify it, build a business on it, keep your changes private. It is the same licence a great deal of ordinary open-source software uses, and applying it to a frontier-scale model is a deliberate and unusual choice.
Two consequences worth caring about:
This is where enthusiasm meets hardware.
"Free to download" and "free to use" are different claims. To serve V4-Pro you need all 1.6T parameters in fast memory. Even heavily quantised that is a multi-GPU server, not a gaming PC and not a laptop. V4-Flash at 284B is far more approachable and still well beyond consumer hardware.
So who is this actually for?
If you are an individual, the realistic route is DeepSeek's own app and API, or a third-party host. You get the model without the rack.
Honestly? Not much this week, and quite a lot over a year.
Open weights at frontier capability do three things that reach ordinary users indirectly. They drag prices down across the market. They make lock-in less dangerous, because a credible alternative exists that nobody can withdraw. And they widen who can build, which means more tools, including from people who could never have afforded to train anything.
What it does not change is the thing that actually determines your results. A better model does not rescue a vague prompt; it produces a more articulate version of the wrong answer. The skills that decide whether AI is useful to you — describing a task precisely, giving enough context, setting constraints, and checking output you cannot verify at a glance — are model-independent. They were true of V3, they are true of V4, and they will be true of whatever ships next.
That is the case for spending your time on the durable half. Iro's paths cover how to structure a prompt and how to catch a confident wrong answer, which pay off across every model in this post and every one after it.
Iro AI turns ideas like the ones in this post into 5-minute exercises with feedback. Free tier, Pro from $4.17 a month ($49.99 a year, 7-day free trial).
DeepSeek V4 comes in two sizes. V4-Pro has 1.6 trillion total parameters with about 49 billion active per token, and V4-Flash has 284 billion total with about 13 billion active. Both are mixture-of-experts models, so only a fraction of the network runs for any given token even though the whole model must be held in memory.
The weights are released under the MIT licence on Hugging Face, which is unusually permissive: you can download, run, modify and commercially use them with essentially no restrictions. Purists note that open weights are not the same as fully open source, since training data and code are not necessarily included, but MIT weights are far more permissive than most licences marketed as open.
Almost certainly not. Mixture-of-experts means only around 49 billion of V4-Pro's parameters are active per token, but all 1.6 trillion must still be resident in memory, which requires a multi-GPU server rather than a desktop. V4-Flash at 284 billion is more approachable but still beyond consumer hardware. Most people will reach V4 through DeepSeek's own app and API or a third-party inference provider.
Both V4-Pro and V4-Flash support a 1 million token context window, according to their Hugging Face model cards.
Only partly. A stronger model raises the ceiling but does not fix a vague request — it tends to produce a more fluent version of the wrong answer. The factors that most affect your results are describing the task precisely, supplying enough context, stating constraints, and verifying output you cannot check at a glance. Those skills transfer across every model.