What Is DeepSeek-V3.2? Open-Weight Model (2026)

DeepSeek-V3.2 is an open-weight DeepSeek model you can self-host or reach through third-party endpoints; in our OpenRouter test it served a 27-token prompt and a 72-token reply in 6.05 seconds for $0.00013608, at listed rates of $0.229 input and $0.343 output per million tokens. That single call is the anchor for everything below.

What is DeepSeek-V3.2? (verified 2026-07)

DeepSeek-V3.2 is an open-weight large language model that DeepSeek publishes for self-hosting and third-party serving, distinct from the closed V4 line the company sells only through its own API. The "open-weight" label is the whole point: the weights are downloadable, so independent hosts can serve the model and set their own rates. That is why you find V3.2 on aggregator endpoints instead of exclusively at api.deepseek.com.

The distribution model changes how you shop for it. A closed model carries one price, set by its vendor. An open-weight model like V3.2 carries as many prices as it has hosts, and the number you actually pay depends entirely on which endpoint you route to. So the rest of this piece is concrete: the numbers we measured on one host, and how they line up against DeepSeek's official V4 line.

Rates as listed on the third-party endpoint we tested (V3.2, Qwen3, GLM-5, MiniMax) and DeepSeek's official docs (V4-Flash), read 2026-07-10. Open-weight rows are one host's snapshot, not a universal price.

We ran DeepSeek-V3.2 on a live endpoint (first-hand evidence)

Documentation tells you a model exists. A live call tells you what it costs and how fast it answers. We sent one short prompt to DeepSeek-V3.2, measured via OpenRouter, on 2026-07-10.

The endpoint served the model as deepseek-v3.2-20251201 and billed the call at $0.00013608. Our prompt was 27 tokens, the reply was 72 tokens, and the round trip took 6.05 seconds. That is a hundredth of a cent for a complete exchange, which tells you the token counts, not the per-million rate, dominate a small call.

Two things stand out. The absolute cost is trivial because the exchange is short. The latency, 6.05 seconds, sits on the slower side of the models we ran that day; the cheapest Qwen3 reply came back in under three seconds. For a nightly batch job that gap is noise. For an interactive agent it is a delay the user feels on every turn.

One disclosure we attach to every open-weight number: this run went through a third-party router, not DeepSeek's own API, so the served-model string, the billed price, and the latency all belong to that host on that day. We have flagged this run needs_native_reverify, meaning the figures should be re-confirmed against a first-party or self-hosted deployment before anyone budgets against them. Routing changes reality, and the same weights on a different host bill differently.

How V3.2 relates to the V4 line on the official API

If V3.2 is the open-weight release, the V4 line is what DeepSeek sells directly. According to DeepSeek API Docs, the V4-Flash tier lists at $0.14 per million input tokens and $0.28 per million output, which undercuts the $0.229 / $0.343 we saw for V3.2 on the third-party host. The official, newer model is the cheaper one here, an inversion of the usual "open weights are always cheaper" assumption.

So why would official V4-Flash beat an open-weight V3.2 on price? According to DeepSeek API Docs, the first-party rate card also carries a prompt-cache discount that third-party hosts do not necessarily pass through, so a workload with a stable prompt prefix can widen the official-API advantage well beyond the headline gap. For the full official rate card and the cache mechanics, see our DeepSeek API pricing hub.

The practical read is a clean split. Pick V3.2 on an open-weight host when you specifically need the weights themselves, for self-hosting, fine-tuning, or on-prem compliance. If you just want DeepSeek quality at the lowest per-token price on a hosted API, the official V4 line is the cheaper default. According to DeepSeek API Docs, those official rates apply at api.deepseek.com, which is the endpoint we recommend budgeting against for the closed line.

Where V3.2 sits among open-weight releases

V3.2 arrived in the same window as a wave of open-weight Chinese models, and the price spread across that set is wide. On the endpoint we tested, Qwen3-235B-A22B listed the lowest output rate at $0.10 per million, an order of magnitude under V3.2's $0.343, while GLM-5 sat at the top of the group at $1.92 output. V3.2's $0.229 / $0.343 lands in the mid-tier: not the cheapest option on paper, but competitive for a general-purpose model.

On the coding-specialist side, other open-weight releases such as Moonshot's Kimi-Dev-72B report strong SWE-bench Verified scores in public benchmark write-ups, around 60% by those reports. We did not call that model here, so we treat its figures as sourced research rather than first-hand evidence. The takeaway for anyone comparing open-weight Chinese models is that "open-weight" alone tells you almost nothing about price; the host and the specific model set the number.

FAQ

Is DeepSeek-V3.2 open source? It is open-weight: the model weights are published so third parties can host it and you can self-host, which is why it appears on aggregator endpoints rather than only at api.deepseek.com. Whether the full training recipe qualifies as "open source" is a stricter question; the fact buyers care about is that the weights are downloadable and independently servable.

How much does DeepSeek-V3.2 cost? On the third-party endpoint we tested on 2026-07-10, V3.2 listed at $0.229 per million input tokens and $0.343 per million output. Our single live call, 27 prompt tokens and 72 completion tokens, billed $0.00013608. Prices vary by host, so treat these as one host's dated snapshot, not a universal rate.

Is DeepSeek-V3.2 cheaper than the official V4 line? Not in our reading. The official V4-Flash tier lists at $0.14 / $0.28 per million, below the $0.229 / $0.343 we saw for V3.2 on a third-party host, and the official API adds a prompt-cache discount on top. Choose V3.2 for the open weights, not to beat the official API on price.

How fast is DeepSeek-V3.2? In our test the round trip took 6.05 seconds for a 72-token reply, measured via a third-party router. That was on the slower side of the open-weight models we ran the same day. Latency depends heavily on the host and its load, so re-verify against your own endpoint before relying on it.

Can I trust these numbers? They are first-hand for the run we made, but routed through a third-party host rather than DeepSeek's own API, so we flagged the run for native re-verification. Re-confirm the price, the served-model string, and the latency against a first-party or self-hosted deployment before you budget.

This is part of the open-source AI agents hub, where we track newly released open-weight models and the frameworks that call them. For the official DeepSeek rate card, see the DeepSeek API pricing hub.

Author: Kevin Fan, Customer Success Manager at China LLM Directory, specializing in Chinese LLM ecosystem pricing and model benchmarking. Last verified: 2026-07-10.