DeepSeek V3.2
DeepSeek · released Dec 1, 2025 · deepseek-ai/DeepSeek-V3.2
- Type
- Open weightsMIT License
- Params
- 685B
- Context
- 164K
37B active per word · about 123K words of context
Our take
Written Aug 2, 2026DeepSeek V3.2 is a large downloadable text model with a permissive MIT licence and a 163,840-token request limit. Only 37 billion of its 685.4 billion total parameters are active per token, keeping inference costs down while measured coding performance is the clear peak in its benchmark profile.
Pick this for production coding workloads where its measured coding score is the highest point in its own benchmark set. Use it for cost-sensitive inference at scale, or for long-context document processing with a permissive licence that allows commercial modification and redistribution. Skip it if you need image, video or audio input, or if creative writing quality matters more than code — that is the weakest measured band in its profile.
The case for it
- Coding is the clear peak in its benchmark profile, with an 83.59-point gap over its own creative-writing score.
- Only 37 billion active parameters from 685.4 billion total — about one in nineteen parameters fires per token.
- MIT licence permits commercial use, modification and redistribution without copyleft requirements.
- Ten offers from eight providers, with the fastest measured throughput on the cheapest host.
The case against it
- Creative writing is the weakest measured band in its own benchmark set, well below its coding peak.
- Throughput is unmeasured on four of ten offers, including multiple entries from the same provider.
- Text-to-text only; no image, video or audio input.
How good is it?
IntelligencePuzzles, maths, exam questions
Arena Text (overall)56th of 143 · 1425
Also on this board: 1422.8 via Thinking (Aug 2, 2026). Read the pair, not the higher one.
CodingWriting and fixing code on its own
Arena Coding56th of 143 · 1469.6
Also on this board: 1475.3 via Thinking (Aug 2, 2026). Read the pair, not the higher one.
AgenticPlanning, calling tools, staying on task
DeepSeek V3.2 is not on Arena Agent (IPS), which is where the rating would come from, so there is no rating here. It is on SWE-bench Verified, in 11th of 39 with 70.
WritingWe do not rate this
Two boards come close and neither tests writing: Arena Creative Writing asks people which of two replies they prefer, and LiveBench Language tests whether a model understood a passage. So we show where DeepSeek V3.2 placed and give it no mark out of five.
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, which is why they get no rating.
Every published score for this model8 scoresEvery figure we hold, from 8 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?
- Fits in memory
- weights load entirely on the card
- Spills to system RAM
- some weights offload; much slower
- Too large
- will not load even with offload
- est
- size is calculated; the verdict could change by 10%
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.
All 3 sizes here are calculated, not measured. We hold no measured file for this model, so each size comes from the parameter count and every verdict above inherits that uncertainty.
Check against your own machine → · All 71 devices, with every size →
Or rent it from someone else
Cheapest of 19 live listings. Picked at the widest standard context we hold, within one quantisation slice, so the numbers beside it are a price one host actually charges.
- per 1M tokens
- $0.27 in / $0.40 out
- Context served
- 164K
- Throughput
- Not measured
| Provider | In / out per 1M tokens | Context | Throughput | Trains on prompts | Logs prompts | Zero retention |
|---|---|---|---|---|---|---|
| Baidufp8 | $0.21 / $0.31 | 131K | 26 tok/s | No | Yesunknown period | Unknown |
| StreamLakefp8 | $0.21 / $0.32 | 128K | 20 tok/s | No | Yesunknown period | Unknown |
| DeepInfrafp4 | $0.26 / $0.38 | 164K | 12 tok/s | No | No | Confirmed |
| AtlasCloudfp8 | $0.26 / $0.38 | 164K | 21 tok/s | No | Yesunknown period | Unknown |
| DeepInfrafp4 | $0.26 / $0.38 | 164K | not measured | Unknown | Unknown | Unknown |
| Novita AIfp8 | $0.27 / $0.40 | 164K | 22 tok/s | No | No | Confirmed |
| Novita AI | $0.27 / $0.40 | 164K | not measured | Unknown | Unknown | Unknown |
| OpenRouter | $0.27 / $0.40 | 164K | not measured | Unknown | Unknown | Unknown |
| SiliconFlowfp8 | $0.26 / $0.42 | 164K | 25 tok/s | No | No | Confirmed |
| GMICloudfp8 | $0.29 / $0.43 | 164K | 28 tok/s | No | Yesunknown period | Unknown |
| Venice AI | $0.33 / $0.48 | 160K | 7 tok/s | No | No | Confirmed |
| Phala | $1.00 / $1.00 | 164K | 7 tok/s | No | No | Confirmed |
| Alibaba Cloudfp8 | $0.37 / $1.11 | 131K | 40 tok/s | No | Yesunknown period | Unknown |
| Alibaba Cloud | $0.37 / $1.11 | 131K | 34 tok/s | No | Yesunknown period | Unknown |
| DigitalOcean Gradient | $0.42 / $1.36 | 164K | 40 tok/s | No | No | Confirmed |
| Friendli | $0.50 / $1.50 | 164K | 26 tok/s | No | Yesunknown period | Unknown |
| Google Vertex AI | $0.56 / $1.68 | 164K | 14 tok/s | No | No | Confirmed |
| SambaNova | $3.00 / $4.50 | 33K | 38 tok/s | No | No | Confirmed |
| SambaNova | $3.00 / $4.50 | 33K | not measured | Unknown | Unknown | Unknown |
Baidu
DeepInfra
OpenRouter
DigitalOcean Gradient
Google Vertex AI
SambaNovaAcross the 19 listings we hold: 15 say they do not train on prompts, 0 say they do and 4 do not say. 8 appear in the zero-retention registry we check; the rest are unknown to us rather than confirmed either way.
What each host's API supports
From the parameter list each endpoint publishes. Streaming is omitted: nothing we hold reports it, for any model.
- ✓
- Supported
- ✗
- Not supported
- Not published
- host gave no parameter list
| Provider | Tool calling | JSON output | Strict schema |
|---|---|---|---|
| Baidufp8 | ✗ | ✓ | ✓ |
| StreamLakefp8 | ✓ | ✓ | ✓ |
| DeepInfrafp4 | ✓ | ✓ | ✓ |
| AtlasCloudfp8 | ✓ | ✓ | ✓ |
| DeepInfrafp4 | |||
| Novita AIfp8 | ✓ | ✗ | ✗ |
| Novita AI | |||
| OpenRouter | ✓ | ✓ | ✓ |
| SiliconFlowfp8 | ✓ | ✓ | ✓ |
| GMICloudfp8 | ✓ | ✗ | ✗ |
| Venice AI | ✓ | ✓ | ✓ |
| Phala | ✓ | ✓ | ✓ |
| Alibaba Cloudfp8 | ✓ | ✓ | ✓ |
| Alibaba Cloud | ✓ | ✓ | ✓ |
| DigitalOcean Gradient | ✗ | ✗ | ✗ |
| Friendli | ✓ | ✓ | ✓ |
| Google Vertex AI | ✓ | ✓ | ✓ |
| SambaNova | ✗ | ✗ | ✗ |
| SambaNova |
Tool calling: 13 of 19 listings say yes, 3 say no, 3 publish no parameter list. JSON output: 12 of 19 listings say yes, 4 say no, 3 publish no parameter list. Strict schema: 12 of 19 listings say yes, 4 say no, 3 publish no parameter list.
Models people weigh against DeepSeek V3.2
When we formed this view
Dates behind this page
Prices last checked 6h ago
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.
- 3 of 19 listings publish no parameter list, so what their API accepts is unknown to us.
- Nothing we hold says whether an endpoint streams, so we do not show it either way.
- 4 of 19 listings do not say whether they train on prompts.
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
- Architecture
- Mixture of experts
- Modality record
- text->text
- Catalogue slug
- deepseek-deepseek-v3-2