DeepSeek V3.2 Exp
DeepSeek · released Sep 29, 2025 · deepseek-ai/DeepSeek-V3.2-Exp
- Type
- Open weightsMIT License
- Params
- 685B
- Context
- 164K
37B active per word · about 123K words of context
Our take
The case for it
- 37 billion of its 685.4 billion parameters work per token, so memory in use is closer to a mid-size model than the total size suggests.
- The licence allows commercial use, changes and redistribution (MIT).
- The request capacity takes long reports or stacks of documents without splitting them first, though reliable recall across all of it is unverified in our data.
The case against it
- 78th of 168 on Arena Text (overall) as of 25 Sep 2026, a board that records which answer people preferred rather than whether it was correct; nothing supplied measures coding or reasoning, so those need a trial on work you can check yourself.
- 84th of 95 on Arena Code (WebDev) as of 25 Sep 2026, among its weaker measured categories; that board records human votes on web-app building tasks rather than whether the code works.
How good is it?
EverydayGeneral questions and everyday reasoning
Arena Text (overall)78th of 168 · 1422
Also on this board: 1425 (Sep 25, 2026). Read the pair, not the higher one.
CodingWriting and fixing code on its own
Arena Coding77th of 168 · 1467
Also on this board: 1475 (Sep 25, 2026). Read the pair, not the higher one.
AgenticPlanning, calling tools, staying on task
Not yet scored on Arena Agent.
WritingDrafting and rewriting prose
Arena Creative Writing55th of 168 · 1409
Arena Creative Writing is the only board that has scored it for this.
Also on this board: 1391 (Sep 25, 2026). Read the pair, not the higher one.
These tests check whether a model follows instructions — a precondition for all the work above, but not a measure of how well that work is done.
Every published score for this model7 scoresEvery figure we hold, from 7 boards, with who ran it and a link to the source — including the boards no rating above is built on.
Can you run it yourself?
GeForce RTX 4090 · 24 GB
Too large for this card. The weights do not fit even with part of them offloaded to system memory.
Apple M1 Pro (16-core GPU) · 32 GB
Too large for this card. The weights do not fit even with part of them offloaded to system memory.
Radeon RX 7900 XT · 20 GB
Too large for this card. The weights do not fit even with part of them offloaded to system memory.
Memory use by level
Against a 24 GB card.
What is quantisation? →This model on every device we track71 devicesThe Q4 build most people download, on each device: what the weights come to, how much context the memory leaves, and whether it runs. Smallest device that runs it first.
Check against your own machine → · Where to rent it hosted →
Or rent it from someone else
Prices checked between 1 hour and 4 days ago — each listing carries its own date.
- per 1M tokens
- $0.27 in / $0.41 out
- Context served
- 164K
- Throughput
- Not measured
| Provider | In / out per 1M tokens | Context | Throughput | Trains on prompts | Logs prompts | Zero retention |
|---|---|---|---|---|---|---|
| OpenRouterOpenRouter's own listing | $0.27 / $0.41checked 1 hour ago | 164K | not measured | Unknown | Unknown | Unknown |
| Novita AIDirect | $0.27 / $0.41checked 4 days ago | 164K | not measured | Unknown | Unknown | Unknown |
| SiliconFlowfp8Through OpenRouter | $0.27 / $0.41checked 1 hour ago | 164K147K max reply | 13 tok/s | No | No | Confirmed |
OpenRouterAcross the 3 listings we hold: 1 says it does not train on prompts, 0 say they do and 2 do not say. 1 appears in the zero-retention registry we check; the rest are unknown to us.
What each host's API supports
From the parameter list each endpoint publishes. Streaming is omitted: nothing we hold reports it, for any model.
| Provider | Tool calling | JSON output | Strict schema |
|---|---|---|---|
| OpenRouterOpenRouter's own listing | ✓ | ✓ | ✓ |
| Novita AIDirect | |||
| SiliconFlowfp8Through OpenRouter | ✓ | ✓ | ✓ |
Tool calling: 2 of 3 listings say yes, 1 publishes no parameter list. JSON output: 2 of 3 listings say yes, 1 publishes no parameter list. Strict schema: 2 of 3 listings say yes, 1 publishes no parameter list.
When we formed this view
Recent changes
What moved
input −50% ($0.270 → $0.134 per 1M tokens), output −51% ($0.41 → $0.20 per 1M tokens), cache read −75% ($0.270 → $0.067 per 1M tokens)What moved
first indexed by our pipelineEach date is the day we first saw the change, or the day the maker announced it.
What we do not know about this model yet
- We hold no measured file for it, so all 3 sizes on this page are calculated from the parameter count.
- 1 of 3 listings publishes no parameter list, so what its API accepts is unknown to us.
- Nothing we hold says whether an endpoint streams, so we do not show it either way.
- 2 of 3 listings do not say whether they train on prompts.
- We hold no batch or off-peak rate for any of its listings.
- We hold a decode speed for it, but no prompt-processing (prefill) figure, so how long the input side of a job takes is unknown to us.
Licence and identifiers
What the licence allowsMIT License, what it allows commercially, and the identifiers you need to pull this model — its Hugging Face repo, our slug and a machine-readable card.
Licence
MIT License
Fully permissive: do anything with attribution. No patent grant, unlike Apache-2.0.
Identifiers
- Hugging Face
- deepseek-ai/DeepSeek-V3.2-Exp
- Architecture
- Mixture of experts
- Takes in, gives back
- Text in, text out
- Catalogue slug
- deepseek-deepseek-v3-2-exp