Nex-N2.5 Pro vs DeepSeek V4 Pro: Only One Set of These Open Weights Is Actually Open

Here is the family secret that makes this matchup stranger than it looks: Nex-AGI's own Nex-N2.5 Max, the top tier of the family that includes Nex-N2.5 Pro, is itself post-trained from DeepSeek-V4-Pro-Base. Coverage of the September 8 release notes that openly — Mini and Pro are built from Qwen3.5 variants, while the text-only 1.6T Max stands on the open base DeepSeek released in August. So when Nex-N2.5 Pro is compared with DeepSeek V4 Pro, the reader is watching one open-model project borrow the other's foundation for its flagship tier and then try to out-position the donor in the tier below. That is not an insult; it is how the open-weights economy works, and it is the most honest possible frame for a comparison where both models are open-licensed and only one of them can actually be downloaded today.

The asymmetry is the whole story. DeepSeek V4 Pro, the 0813 build that went GA on August 13, has MIT-licensed weights sitting on Hugging Face right now — a large mixture-of-experts model of roughly 1.6 trillion total parameters with about 49 billion active, a 1M-token context, a 384K output ceiling, and an official API behind it. Nex-N2.5 Pro, announced September 8, is a 397-billion-parameter sparse MoE with about 17 billion active, multimodal where DeepSeek is text-only, and aimed at computer-use agents rather than generalist tool use. Its Hugging Face card still banners that the weights are coming soon under Apache-2.0, with no date, while only free hosted endpoints serve it. Same philosophical commitment to open weights. Completely different state of delivery.

Two licenses, one set of weights actually released

The license contrast matters less than the delivery contrast, but it sets the terms. DeepSeek V4 Pro is MIT — you can download the shards today, serve them yourself, fine-tune them, and build a business on them without asking anyone's permission, and because the weights are out, no single host controls the price. Nex-N2.5 Pro is promised under Apache-2.0, the equally permissive license the alliance chose, but the banner on the card is unambiguous: the weights are not downloadable, and no date is attached. For a team whose requirement is "the model must be ours," this matchup is currently not a contest — DeepSeek V4 Pro is the only candidate you can actually own. The practical caveat cuts both ways: self-hosting a ~1.6T MoE is serious infrastructure, so for most teams the MIT license matters less as a hosting plan than as a guarantee against lock-in and a floor under price. Nex-N2.5 Pro, when its shards land, will be the more self-hostable design — 17B active on an eight-H100 node is a far smaller box than DeepSeek's reference footprint.

The rows where the two differ

Side by side, the two models barely overlap in what they are for:

• Released — Nex-N2.5 Pro: September 8, 2026 (hosted endpoints live, weights pending). DeepSeek V4 Pro: August 13, 2026 (0813 GA build), weights live.

• Scale — Nex-N2.5 Pro: 397B total / ~17B active. DeepSeek V4 Pro: ~1.6T total / ~49B active.

• Modality — Nex-N2.5 Pro: text and image input, text output; built for screen-reading computer use. DeepSeek V4 Pro: text-only in its public API — no native vision, image or audio input.

• Context / output — Nex-N2.5 Pro: 262,144-token context. DeepSeek V4 Pro: 1M-token context, up to 384K output tokens.

• Reasoning control — Nex-N2.5 Pro: none / medium (adaptive, default) / high. DeepSeek V4 Pro: low / high / max reasoning effort on the same model string.

• Weights — Nex-N2.5 Pro: Apache-2.0, coming soon. DeepSeek V4 Pro: MIT, downloadable now.

• API surface — Nex-N2.5 Pro: OpenAI-compatible chat calls on free hosted endpoints. DeepSeek V4 Pro: native OpenAI Responses API plus Anthropic-compatible endpoints, tool calls and JSON output.

Two genuinely different products share the word "open." Nex-N2.5 Pro is built to look at a screen and drive it; DeepSeek V4 Pro is built to think, write code, and call tools — the difference between an operator and an analyst.

What is measured, and what is borrowed

The benchmark picture is as lopsided as the delivery picture. DeepSeek V4 Pro has a measured, independent footprint: an entry on the Artificial Analysis Intelligence Index, weeks of press and production traffic since the 0813 GA, and vendor coding figures — Terminal-Bench 2.1 at 87.9 and DeepSWE at 62.7 on DeepSeek's own reporting — that at least sit on top of a model anyone can download and re-run. Nex-N2.5 Pro has none of that. Its headline numbers — OSWorld-2 at 56.4, Terminal-Bench 2.1 at 82.7, SWE-Bench Pro at 61.2, BrowseComp at 89.7 — are all Nex-AGI's own measurements through its NexCUA harness, which the alliance says it will open-source but has not yet. In the first 48 hours after launch, no independent laboratory had run Nex-N2.5 Pro, because no one outside the alliance can.

Read the two sets of claims side by side and the honest conclusion is not "DeepSeek is smarter." It is "DeepSeek's numbers are checkable and Nex-N2.5 Pro's are not." Where the two vendor sheets do overlap — Terminal-Bench 2.1 — DeepSeek claims 87.9 against Nex-N2.5 Pro's 82.7, a five-point gap in the incumbent's favor that is exactly the direction a cautious reader should expect: the model that has shipped measures higher on the shared row than the one that has not.

The price: a real number, a scheduled number, and no number

DeepSeek V4 Pro's pricing is not one number, and that is its own story. At the 0813 GA the official API listed $0.435 per million input and $0.87 per million output; on August 16 DeepSeek moved the V4 family to a peak/off-peak schedule, so the official API now charges $0.66 input / $1.98 output off-peak and $1.32 / $3.96 at peak, with cached input at $0.022 off-peak and $0.044 at peak. Depending on when you call DeepSeek directly, the same model costs between roughly 50% more and four and a half times the launch rate. Route matters too: the OrcaRouter model page for DeepSeek V4 Pro currently lists a flat $0.44 input / $0.88 output per million — the provider's rate passed through with no markup — which undercuts even the off-peak official schedule. That flat route is the price to beat if you are comparing cost per token, and it is the kind of number that only exists because a routing layer shows you the provider's real rate rather than a repriced one.

Nex-N2.5 Pro has no price to put next to any of that. Its hosted tier is free, and Nex-AGI has published no paid per-token rate for the family, so the model's economics are simply unknown — the free endpoints prove the build works, nothing more. If the weights land and a provider starts serving Nex-N2.5 Pro at a list price, the honest comparison becomes possible; until then, "free" is a preview mechanic, not a market signal.

Two different kinds of agent

Choose between these two and you are choosing between two jobs, not two scores. Nex-N2.5 Pro is for the agent that operates a computer: it reads a screen, moves a cursor, edits a file, checks the visual result, and keeps going — the demonstrated 468-action session in Pokemon Platinum is the shape of that ambition. DeepSeek V4 Pro is for the agent that operates an API: it reasons, writes multi-file code, calls tools, and produces long structured output, all through the text channel that its 384K output ceiling and 1M context exist to serve. A team building browser automation should be watching Nex-N2.5 Pro. A team building a coding or knowledge-work agent should be comparing DeepSeek V4 Pro against the closed frontier, and Nex-N2.5 Pro is not yet a candidate for that job — it cannot take the image out of the loop, and its coding claims are both lower and less verified than the model it is being compared with.

The verdict: wait on one, build on the other

DeepSeek V4 Pro is the only rational build target in this matchup today. Its weights are downloadable, its price is known — and cheaper through the flat OrcaRouter route than on DeepSeek's own peak schedule — and its claims can be verified by anyone in an afternoon. It is also the natural place to start measuring what open computer-use will cost: route its coding-agent traffic through OrcaRouter with automatic failover around the official API's ~500-concurrency cap, and you have a working open-weights baseline at a real price. Nex-N2.5 Pro is the more interesting long-term bet precisely because it is the more specialized design — a 17B-active multimodal operator that could undercut the closed computer-use leaders the way DeepSeek undercut the closed generalists — but an unverifiable free endpoint is not a foundation. When the Apache-2.0 shards land and an independent harness runs that OSWorld-2 number, this matchup changes overnight. Until then, it is a comparison between a model you can own and a model you can only watch.