DeepSeek V3.2
DeepSeekhuggingface.co/deepseek-ai/DeepSeek-V3.2
DeepSeek-V3.2 is a large language model designed to harmonize high computational efficiency with strong reasoning and agentic tool-use performance. It introduces DeepSeek Sparse Attention (DSA), a fine-grained sparse attention mechanism...
Updated 2 h ago · first seen 11 Sept 2026
model_01M294WW8H2JQJXMSPY7QTDVWF
- Parameters
- 685.4B
- T2 · 5 h ago
- Context
- 163.8K tokens
- T2 · 2 h ago
- Released
- 1 Dec 2025
- T2 · 5 h ago
- License
- mit
- T2 · 4 h ago
Specification
- Release date
- 1 Dec 2025
Source:OpenRouter public model & pricing listingT2observed 5 h agomedium
- Status
- deprecated
Source:Artificial AnalysisT2observed 3 h agomedium
- Openness
- open-weights
Source:Hugging Face Hub (public pages, model cards, papers)T2observed 5 h agomedium
- License
- mit
Source:Hugging Face Hub (public pages, model cards, papers)T2observed 4 h agomedium
- Architecture
- DeepseekV32ForCausalLM
Source:Hugging Face Hub (public pages, model cards, papers)T2observed 4 h agomedium
- Parameters
- 685.4B
Source:Hugging Face Hub (public pages, model cards, papers)T2observed 5 h agomedium
- Context window
- 163.8K tokens
Source:OpenRouter public model & pricing listingT2observed 2 h agomedium
- Max output
- 65.5K tokens
Source:OpenRouter public model & pricing listingT2observed 5 h agomedium
- Modalities
- text
Source:OpenRouter public model & pricing listingT2observed 5 h agomedium
- Input modalities
- text
Source:OpenRouter public model & pricing listingT2observed 5 h agomedium
- Output modalities
- text
Source:OpenRouter public model & pricing listingT2observed 5 h agomedium
- Tokenizer
- DeepSeek
Source:OpenRouter public model & pricing listingT2observed 5 h agomedium
- Base model
- deepseek-ai/DeepSeek-V3.2-Exp-Base
Source:Hugging Face Hub (public pages, model cards, papers)T2observed 4 h agomedium
- Quantization format
- fp8
Source:Hugging Face Hub (public pages, model cards, papers)T2observed 4 h agomedium
- File size
- 689.5 GB
Source:Hugging Face Hub (public pages, model cards, papers)T2observed 4 h agomedium
- Hugging Face repo
- deepseek-ai/DeepSeek-V3.2
Source:OpenRouter public model & pricing listingT2observed 5 h agomedium
- Pipeline tag
- text-generation
Source:Hugging Face Hub (public pages, model cards, papers)T2observed 5 h agomedium
- Model card
Source:Hugging Face Hub (public pages, model cards, papers)T2observed 5 h agomedium
- Downloads
- 1,732,352
Source:Hugging Face Hub (public pages, model cards, papers)T2observed 3 h agomedium
- Likes
- 1,477
Source:Hugging Face Hub (public pages, model cards, papers)T2observed 3 h agomedium
- Gated
- No
Source:Hugging Face Hub (public pages, model cards, papers)T2observed 4 h agomedium
- Hf inference providers
- deepinfra, featherless-ai, novita
Source:Hugging Face Hub (public pages, model cards, papers)T2observed 4 h agomedium
- Is quantized
- Yes
Source:Hugging Face Hub (public pages, model cards, papers)T2observed 4 h agomedium
- Last modified
- 2025-12-01T11:04:59+00:00
Source:Hugging Face Hub (public pages, model cards, papers)T2observed 5 h agomedium
- Library name
- transformers
Source:Hugging Face Hub (public pages, model cards, papers)T2observed 4 h agomedium
- Downloads all time
- 23,362,767
Source:Hugging Face Hub (public pages, model cards, papers)T2observed 3 h agomedium
- Model type
- deepseek_v32
Source:Hugging Face Hub (public pages, model cards, papers)T2observed 4 h agomedium
- Openrouter id
- deepseek/deepseek-v3.2
Source:OpenRouter public model & pricing listingT2observed 5 h agomedium
- Supported parameters
- frequency_penalty, include_reasoning, logit_bias, logprobs, max_tokens, min_p, presence_penalty, reasoning, repetition_penalty, response_format, seed, stop, structured_outputs, temperature, tool_choice, tools, top_k, top_logprobs, top_p
Source:OpenRouter public model & pricing listingT2observed 5 h agomedium
- Tags
- transformers, safetensors, deepseek_v32, text-generation, fp8
Source:Hugging Face Hub (public pages, model cards, papers)T2observed 4 h agomedium
- Weights dtype
- BF16, F32, F8_E4M3
Source:Hugging Face Hub (public pages, model cards, papers)T2observed 4 h agomedium
Each value shows its source, tier and observation time. Conflicting claims are kept side by side and flagged — never averaged. How AI Atlas records facts →
Provenance
Attributed facts
35
Source tiers
T235
Freshest observation
2 h ago
Conflicts
None
Modalities
- Modalities
- text
- Input
- text
- Output
- text
Capabilities
Tool calling
Yes
OpenRouter public model & pricing listing · T2
Structured output
Yes
OpenRouter public model & pricing listing · T2
Reasoning
Yes
OpenRouter public model & pricing listing · T2
Vision
Unavailable
Audio
Unavailable
Fine-tuning available
Unavailable
- Context window
- 163.8K tokens
Source:OpenRouter public model & pricing listingT2observed 2 h agomedium
- Max output
- 65.5K tokens
Source:OpenRouter public model & pricing listingT2observed 5 h agomedium
- Tokenizer
- DeepSeek
Source:OpenRouter public model & pricing listingT2observed 5 h agomedium
| Benchmark | Score | Metric | Config | Evaluated | Source |
|---|---|---|---|---|---|
| τ²-bench | 90.64% | pass^1 | aa_slug=deepseek-v3-2-reasoning · evaluator=Artificial Analysis · index_version=4.3 | — | artificialanalysis.aiT2 |
| Artificial Analysis Intelligence Index | 21.49 | index | aa_slug=deepseek-v3-2-reasoning · version=4.3 · estimated=true | — | artificialanalysis.aiT2 |
| GPQA | 84.04% | accuracy | aa_slug=deepseek-v3-2-reasoning · variant=GPQA Diamond · evaluator=Artificial Analysis · index_version=4.3 | — | artificialanalysis.aiT2 |
| Humanity's Last Exam | 24.56% | accuracy | aa_slug=deepseek-v3-2-reasoning · evaluator=Artificial Analysis · index_version=4.3 | — | artificialanalysis.aiT2 |
| IFBench | 60.68% | accuracy | aa_slug=deepseek-v3-2-reasoning · evaluator=Artificial Analysis · index_version=4.3 | — | artificialanalysis.aiT2 |
| SWE-bench Multilingual | 59% | resolved | date=2026-02-13 · board=Multilingual · system=mini-SWE-agent · model_tag=deepseek-v3.2 | 13 Feb 2026 | swebench.comT2 |
| SWE-bench Verified | 70% | resolved | date=2026-02-17 · board=Verified · system=mini-SWE-agent · model_tag=deepseek-v3.2 | 17 Feb 2026 | swebench.comT2 |
| Terminal-Bench | 46.82% | accuracy | aa_slug=deepseek-v3-2-reasoning · variant=v2.1 · evaluator=Artificial Analysis · index_version=4.3 | — | artificialanalysis.aiT2 |
| Terminal-Bench | 35.61% | accuracy | aa_slug=deepseek-v3-2-reasoning · variant=hard · evaluator=Artificial Analysis · index_version=4.3 | — | artificialanalysis.aiT2 |
Scores are reported as published, with their evaluation configuration (harness, prompting, judge). Results with different configs are not directly comparable — see methodology.
Current prices
| Provider | Input / 1M | Output / 1M | Cached in | Batch in / out | Context | Observed | Source |
|---|---|---|---|---|---|---|---|
| DeepSeek APIdeepseek/deepseek-v3.2 | $0.269 | $0.40 | $0.135 | — | 163.8K | 37 min ago | openrouter.aiT2 |
USD per 1M tokens as published by each provider (USD). Rows are append-only: every change is kept in the history below.
Available through 1
Price history
Input price · USD / 1M tokens
Output price · USD / 1M tokens
- DeepSeek API
Memory need = bytes per parameter (4-bit ≈ 0.5 × 1.15 overhead, 8-bit 1.0, fp16 2.0) + a KV-cache allowance. Not a measurement.
| Hardware | Quantization | Memory | Est. need | Fits |
|---|---|---|---|---|
| NVIDIA DGX B200 | 4bit | 1,440 GB | 394.6 GB | Yes |
| Apple M3 Ultra | 4bit | — | 394.6 GB | Yes |
| Mac Studio (Apple M5 Ultra) | 4bit | — | 394.6 GB | Yes |
| AMD Instinct MI325X | 4bit | 256 GB | 394.6 GB | No |
| AMD Instinct MI300X | 4bit | 192 GB | 394.6 GB | No |
| Apple M2 Ultra | 4bit | — | 394.6 GB | No |
| NVIDIA B200 | 4bit | 180 GB | 394.6 GB | No |
| NVIDIA H200 | 4bit | 141 GB | 394.6 GB | No |
| NVIDIA H200 NVL | 4bit | 141 GB | 394.6 GB | No |
| Apple M1 Ultra | 4bit | — | 394.6 GB | No |
| Apple M3 Max | 4bit | — | 394.6 GB | No |
| Apple M4 Max | 4bit | — | 394.6 GB | No |
| Mac Studio (Apple M5 Max) | 4bit | — | 394.6 GB | No |
| MacBook Pro (Apple M5 Max) | 4bit | — | 394.6 GB | No |
| NVIDIA DGX Spark | 4bit | 128 GB | 394.6 GB | No |
| Apple M2 Max | 4bit | — | 394.6 GB | No |
| NVIDIA H100 NVL | 4bit | 94 GB | 394.6 GB | No |
| NVIDIA A100 80GB | 4bit | 80 GB | 394.6 GB | No |
| NVIDIA H100 SXM | 4bit | 80 GB | 394.6 GB | No |
| Apple M1 Max | 4bit | — | 394.6 GB | No |
| Apple M4 Pro | 4bit | — | 394.6 GB | No |
| Mac mini (Apple M5 Pro) | 4bit | — | 394.6 GB | No |
| MacBook Pro (Apple M5 Pro) | 4bit | — | 394.6 GB | No |
| Apple M3 Pro | 4bit | — | 394.6 GB | No |
| Apple M1 Pro | 4bit | — | 394.6 GB | No |
| Apple M2 Pro | 4bit | — | 394.6 GB | No |
| Apple M4 | 4bit | — | 394.6 GB | No |
| iMac (Apple M4) | 4bit | — | 394.6 GB | No |
| Mac mini (Apple M6) | 4bit | — | 394.6 GB | No |
| MacBook Air (Apple M5) | 4bit | — | 394.6 GB | No |
| MacBook Pro (Apple M5) | 4bit | — | 394.6 GB | No |
| NVIDIA GeForce RTX 5090 | 4bit | 32 GB | 394.6 GB | No |
| Apple M2 | 4bit | — | 394.6 GB | No |
| Apple M3 | 4bit | — | 394.6 GB | No |
| NVIDIA GeForce RTX 3090 | 4bit | 24 GB | 394.6 GB | No |
| NVIDIA GeForce RTX 4090 | 4bit | 24 GB | 394.6 GB | No |
| Apple M1 | 4bit | — | 394.6 GB | No |
Ancestors 1
- DeepSeek-V3.2-Exp-BaseDeepSeek
This model
DeepSeek V3.2
685.4B params
Quantizations 2
Descendants 0
None recorded.
Papers 0
No papers linked yet.
Repositories 0
No repositories linked yet.
As of
Rewind the record: see this entity's attributes exactly as AI Atlas knew them on a given day.
Claim history · File size
File sizefile_size_gb1
| Value | Valid from → to | Status | Source | Confidence | Extractor |
|---|---|---|---|---|---|
| 689.5 GB | → current | current | Hugging Face Hub (public pages, model cards, papers)T2 | medium | deterministic |
Claims are temporal and append-only: a new observation closes the previous claim (valid_to) instead of overwriting it. Conflicting claims from different sources are kept side by side and flagged — never averaged. Methodology →
DeepSeek V3.2: context length changed from 128000 to 163840
Context window128K tokens→163.8K tokensopenrouterDeepSeek V3.2 scores 70% on SWE-bench Verified
swebench_leaderboardDeepSeek V3.2 scores 59% on SWE-bench Multilingual
swebench_leaderboardDeepSeek V3.2 scores 35.61% on Terminal-Bench
artificial_analysisDeepSeek V3.2 scores 46.82% on Terminal-Bench
artificial_analysisDeepSeek V3.2 scores 90.64% on τ²-bench
artificial_analysisDeepSeek V3.2 scores 24.56% on Humanity's Last Exam
artificial_analysisDeepSeek V3.2 scores 21.49 on Artificial Analysis Intelligence Index
artificial_analysisDeepSeek V3.2: context length changed from 163840 to 128000
Context window163.8K tokens→128K tokensartificial_analysisDeepSeek API lists DeepSeek V3.2 at $0.269 in / $0.4 out per 1M tokens
openrouter
| Source | Document | Type | Tier | Last observed | Snapshots |
|---|---|---|---|---|---|
| OpenRouter public model & pricing listing | openrouter.ai/api/v1/models | catalogue | T2· Quality secondary | 37 min ago | 4 |
| Hugging Face Hub (public pages, model cards, papers) | huggingface.co/deepseek-ai/DeepSeek-V3.2/raw/main/README.md | model_card | T2· Quality secondary | 2 h ago | 1 |
| Hugging Face Hub (public pages, model cards, papers) | huggingface.co/deepseek-ai/DeepSeek-V3.2 | model_page | T2· Quality secondary | 3 h ago | 2 |
| Hugging Face Hub (public pages, model cards, papers) | huggingface.co/cerebras/DeepSeek-V3.2-REAP-345B-A37B | model_page | T2· Quality secondary | 3 h ago | 3 |
| Hugging Face Hub (public pages, model cards, papers) | huggingface.co/cerebras/DeepSeek-V3.2-REAP-508B-A37B | model_page | T2· Quality secondary | 3 h ago | 3 |
| Hugging Face Hub (public pages, model cards, papers) | huggingface.co/models?author=deepseek-ai&p=0&sort=downloads | listing | T2· Quality secondary | 3 h ago | 3 |
| SWE-bench leaderboards | swebench.com/ | leaderboard | T2· Quality secondary | 3 h ago | 1 |
| Artificial Analysis | artificialanalysis.ai/leaderboards/models | leaderboard | T2· Quality secondary | 3 h ago | 1 |
Tier 1 = official/primary, 2 = quality secondary, 3 = community, 4 = unverified. Every snapshot is archived; see all sources and the methodology.