DeepSeek V3 vs V3.1 vs R1: Which Model to Run Locally (2026 Comparison)

Published October 30, 2025 • Rewritten September 29, 2026

DeepSeek V3 and DeepSeek V3.1 are the same size and shape: a Mixture-of-Experts model with 671B total parameters, 37B active per token and a 128K context window. V3.1 is a retrain that adds a switchable thinking mode and much better tool use. Neither one is a realistic download for a desktop — the standard 4-bit build of either is 404 GB. This page sets out what DeepSeek published about each release, and what it would take to run one. Every number is copied from DeepSeek's model cards, its technical report, or the file listings on Hugging Face and Ollama; this site has not benchmarked these models.

The short answer

Which to pick: if you have the hardware for a 671B model at all, V3.1-Terminus replaces both V3 and R1 — it is one set of weights that answers directly or reasons first, and DeepSeek's own table puts its thinking mode level with R1-0528. If you do not have that hardware, none of the three is the right download; see what replaced them and the small R1 distills below.

Release timeline

Dates are from the DeepSeek API news index and the creation dates of the repositories under deepseek-ai on Hugging Face.

What changed from V3 to V3.1

The V3.1 model card lists three changes, in DeepSeek's words:

• Hybrid thinking mode. "One model supports both thinking mode and non-thinking mode by changing the chat template." In practice the prompt ends with an opening think tag to make the model reason first, or a closing think tag to make it answer directly. V3 has no thinking mode; before V3.1 you ran R1 for that.
• Smarter tool calling. "Through post-training optimization, the model's performance in tool usage and agent tasks has significantly improved." Tool calls are supported in non-thinking mode.
• Higher thinking efficiency. "DeepSeek-V3.1-Think achieves comparable answer quality to DeepSeek-R1-0528, while responding more quickly."

What did not change is the architecture. The card says "the model structure of DeepSeek-V3.1 is the same as DeepSeek-V3", and the parameter counts on Hugging Face are identical. V3.1 is built on a new base checkpoint whose long-context training was extended: the 32K phase grew tenfold to 630B tokens and the 128K phase 3.3 times to 209B tokens. The weights use a different FP8 scale format (UE8M0), which matters only if you run the original weights rather than a GGUF.

The September 2025 Terminus update kept the same capabilities and fixed two complaints: mixed Chinese-English text with occasional stray characters, and agent performance.

On the hosted API, the release note for V3.1 mapped deepseek-chat to non-thinking mode and deepseek-reasoner to thinking mode, both with 128K context.

Published benchmarks

These are DeepSeek's own results from the V3.1 model card. The V3 column is the March 2025 V3-0324 refresh, not the December 2024 original. DeepSeek notes that the SWE-bench rows use its internal code-agent framework, so they are not directly comparable with other vendors' figures.

Two things stand out. The agent rows are where V3.1 moved most: SWE Verified went from 45.4 to 66.0 without thinking. And on maths, non-thinking V3.1 is no better than V3-0324 (AIME 2025 is slightly lower) — the gain there comes entirely from switching thinking on.

The Terminus card reports SWE Verified 68.4, SWE-bench Multilingual 57.8 and Terminal-bench 36.7, with the reasoning scores within a point of V3.1.

For the original December 2024 model, the V3 card reports MMLU 88.5, HumanEval-Mul 82.6, LiveCodeBench 37.6, SWE Verified 42.0, Aider-Polyglot 49.6 and AIME 2024 39.2 for the chat model, and HumanEval 65.2 for the base model.

Hardware requirements

DeepSeek V3 and V3.1 do not fit on any single graphics card. The Ollama build of either is 404 GB, so it needs about 512 GB of system memory; the smallest GGUF anyone publishes is 147.5 GB and needs 192 GB. Only 37B parameters are active per token, but that sets speed, not memory: every one of the 671B weights has to be loaded before the first token comes out.

The arithmetic is parameters × bits per weight ÷ 8. At Q4_K_M that is 671B × 4.82 bits ÷ 8 = 404 GB, which matches the file.

"Memory to plan for" is the file size rounded up to the next common memory configuration, leaving room for the context cache and the operating system. It can be system RAM on a server, unified memory on a Mac, or VRAM summed across several cards. The V3-0324 files from unsloth are within a gigabyte or two of the V3.1 sizes at every level.

What this means for ordinary hardware:

• One 24 GB or 32 GB card: no build fits. The smallest file is over four times the size of an RTX 5090's memory.
• A 128 GB workstation or Mac: no build fits.
• 192–256 GB of memory: only the 1-bit and 2-bit builds load. Expect quality well below the published benchmarks at under 2 bits per weight — that is a general property of extreme quantisation, not something measured here.
• 512 GB of memory: the standard Q4_K_M build loads. That is server territory.

This site has no measured speed for any of these builds, so none is quoted. The MoE VRAM calculator does the same sum for other expert counts, and quantization explained covers what the build names mean.

How to run V3 or V3.1

With Ollama, if you have the memory:

# DeepSeek V3 — Q4_K_M, 404 GB
ollama pull deepseek-v3:671b

# DeepSeek V3.1 (the default tag is the Terminus update) — Q4_K_M, 404 GB
ollama pull deepseek-v3.1:671b

These are the only sizes Ollama lists; there is no small deepseek-v3 tag. Check deepseek-v3 tags and deepseek-v3.1 tags before pulling.

For a smaller quantisation, the GGUF repositories are unsloth/DeepSeek-V3.1-GGUF, bartowski/deepseek-ai_DeepSeek-V3.1-GGUF and unsloth/DeepSeek-V3-0324-GGUF, which load in llama.cpp.

For the original weights on a GPU server, the V3 model card lists SGLang, LMDeploy, TensorRT-LLM and vLLM as supported, on NVIDIA and AMD GPUs. DeepSeek publishes FP8 weights only and provides a script to convert them to BF16. The card also says two things V3.1 users need to get right: the expert-gate bias parameters must be computed in FP32, and FP8 weights must use the UE8M0 scale format.

If what you want is a DeepSeek model that runs on a normal PC, the January 2025 R1 release included distilled models. They are fine-tunes of Qwen2.5 and Llama, not smaller copies of V3, but they do fit:

Setup for those is in the DeepSeek R1 local setup guide; the full-size model is covered in the DeepSeek V3 local setup guide.

Licence

The licence is not the same for every release, which earlier versions of this page got wrong.

• DeepSeek-V3 (December 2024): the code is MIT. The weights are under the DeepSeek Model License, a custom agreement with use-based restrictions. The card states that the V3 series "supports commercial use".
• DeepSeek-V3-0324, V3.1, V3.1-Terminus, R1 and R1-0528: "This repository and the model weights are licensed under the MIT License."

So if the MIT licence matters to you, use V3-0324 or later rather than the original December weights. The R1 distills carry the terms of their base models as well — Apache 2.0 for the Qwen-based ones and Meta's Llama licences for the Llama-based ones.

The $5.6M training cost

The figure is real and it comes from Table 1 of the DeepSeek-V3 technical report: 2,664K GPU hours of pre-training, 119K for context extension and 5K for post-training, 2,788K H800 GPU hours in total. Priced at an assumed rental rate of $2 per GPU hour, that is $5.576M.

The report is explicit about what the number leaves out: it covers "only the official training of DeepSeek-V3, excluding the costs associated with prior research and ablation experiments on architectures, algorithms, or data". It is the cost of one training run at a rental price, not the cost of building the model, and the report makes no comparison with any other company's spending.

The same report describes the architecture. V3 uses Multi-head Latent Attention and the DeepSeekMoE design from V2, adds a load-balancing method that needs no auxiliary loss, and trains with a multi-token prediction objective. It was pre-trained on 14.8 trillion tokens. The published configuration has 61 layers with 256 routed experts and one shared expert per MoE layer, and routes each token to 8 experts.

What replaced V3 and V3.1

Both are superseded. DeepSeek released V3.2 in December 2025 and the V4 series in April 2026. V4-Flash is the one that matters for local use: at 284B total parameters it is well under half the size of V3, with 13B active per token, a 1M-token context and the MIT licence. V4-Pro is 1.6T parameters and larger than anything on this page.

• DeepSeek V4: sizes, context and local setup
• DeepSeek V4 hardware requirements by memory tier
• Best models for a 24 GB GPU, if one card is what you have
• Best local AI coding models

Questions people ask

Is DeepSeek V3.1 better than R1?

On DeepSeek's own table, V3.1 in thinking mode is level with or slightly ahead of R1-0528 on most rows (LiveCodeBench 74.8 against 73.3, AIME 2025 88.4 against 87.5) and slightly behind on GPQA-Diamond (80.1 against 81.0). The practical difference is that V3.1 can also answer without thinking and can call tools, so it covers both jobs with one download.

What is the context window of DeepSeek V3 and V3.1?

128K tokens for both, according to the model cards. Ollama's tag list shows 160K for the Q4_K_M builds, which matches the 163,840 position limit in the model's configuration file; it is the same model.

Why does Hugging Face say 685B when the model is 671B?

The V3 card explains it: 671B is the main model, and the repository also holds 14B of multi-token prediction module weights, which are a training aid.

What SWE-bench Verified score does DeepSeek V3 have?

DeepSeek reports 42.0 for the original V3, 45.4 for V3-0324, 66.0 for V3.1 and 68.4 for V3.1-Terminus. The last three were run in DeepSeek's internal agent framework.

Can I run DeepSeek V3.1 on an RTX 4090 or 5090?

No. The smallest published file is 147.5 GB and those cards have 24 GB and 32 GB. Layers can be offloaded to system RAM, but the whole file still has to fit in memory somewhere.

Sources

• DeepSeek-V3 model card and technical report, arXiv:2412.19437
• DeepSeek-V3-0324 model card
• DeepSeek-V3.1 model card and release note
• DeepSeek-V3.1-Terminus model card
• DeepSeek-R1 model card
• DeepSeek-V4-Pro model card
• Ollama tag lists for deepseek-v3, deepseek-v3.1 and deepseek-r1