DeepSeek R1 Distill: The Six Reasoning Models You Can Run Locally

If you want R1-style reasoning on your own GPU rather than through an API, the DeepSeek R1 Distill family is the practical answer. These are six smaller dense models — 1.5B, 7B, 8B, 14B, 32B and 70B parameters — fine-tuned on about 800,000 samples curated with the full DeepSeek-R1, and announced as open weights alongside R1 on 20 January 2025. They are not new architectures. They are existing Qwen2.5 and Llama 3.1/3.3 checkpoints fine-tuned to reason the way R1 does, and on DeepSeek’s published numbers all six beat GPT-4o-0513 and Claude-3.5-Sonnet-1022 on the AIME 2024 and MATH-500 maths benchmarks, and the 14B and larger also on both coding benchmarks.

This guide walks through each size, the benchmark numbers, where they win and lose, how to run them, and how they fit alongside DeepSeek’s current API models.

What DeepSeek R1 Distill actually is

DeepSeek R1 Distill is a set of six open-weight dense models that inherit the reasoning behaviour of the full DeepSeek-R1 through supervised fine-tuning. According to the DeepSeek-R1 paper (arXiv 2501.12948), DeepSeek fine-tuned open-source Qwen and Llama models directly on the 800k samples curated with DeepSeek-R1, and “for distilled models, we apply only SFT and do not include an RL stage, even though incorporating RL could substantially boost model performance.”

That distinction matters. The full DeepSeek R1 went through reinforcement learning to discover its reasoning patterns. The distilled siblings skipped the RL phase entirely and learned to imitate R1’s outputs. DeepSeek’s model card draws the lesson that the reasoning patterns of larger models can be distilled into smaller models, “resulting in better performance compared to the reasoning patterns discovered through RL on small models.”

The six checkpoints at a glance

DeepSeek announced the family in its DeepSeek-R1 release note of 20 January 2025 (“6 small models fully open-sourced”), and the model card says it open-sourced distilled 1.5B, 7B, 8B, 14B, 32B and 70B checkpoints based on the Qwen2.5 and Llama3 series. Four things vary across the line: the base model, the parameter count, the download size and the licence inherited from the base.

DeepSeek’s DeepSeek-R1 model card says the repository and weights are MIT-licensed and that the Qwen 1.5B/7B/14B/32B variants are derived from the Qwen-2.5 series, “originally licensed under Apache 2.0 License, and now finetuned with 800k samples curated with DeepSeek-R1”; R1-Distill-Llama-8B is derived from Llama3.1-8B-Base under the Llama 3.1 licence, and R1-Distill-Llama-70B from Llama3.3-70B-Instruct under the Llama 3.3 licence. So the base-model licence still applies on top of MIT — the 70B in particular is also governed by Meta’s Llama 3.3 licence. The paper adds that DeepSeek picked Llama-3.3 for the 70B because its reasoning capability is slightly better than Llama-3.1’s. The download sizes are the sum of each repository’s current safetensors files.

Benchmarks: how the distilled models stack up

The distilled models punch above the weight class their parameter count suggests. DeepSeek’s distilled-model evaluation table on the R1 model card (January 2025, DeepSeek-reported) compares them with GPT-4o-0513, Claude-3.5-Sonnet-1022, OpenAI o1-mini and QwQ-32B-Preview:

R1-Distill-Qwen-32B posts the best AIME 2024 pass@1 of the six (72.6) with 94.3 on MATH-500, while R1-Distill-Llama-70B has the best MATH-500 (94.5), GPQA Diamond (65.2) and LiveCodeBench (57.5) scores of the distilled models. On CodeForces, o1-mini’s 1820 rating remains ahead of every distilled model.

The 32B is the result most readers remember: DeepSeek’s model card says DeepSeek-R1-Distill-Qwen-32B “outperforms OpenAI-o1-mini across various benchmarks, achieving new state-of-the-art results for dense models”, and the R1 release note described the 32B and 70B as “on par with OpenAI-o1-mini” — both statements from January 2025.

One caveat the card itself flags: these benchmarks are sampled, not deterministic. DeepSeek set the maximum generation length to 32,768 tokens and, for benchmarks requiring sampling, used a temperature of 0.6, a top-p value of 0.95, and 64 responses per query to estimate pass@1. The paper explains why: greedy decoding of long-output reasoning models gave higher repetition rates and significant variability across checkpoints, so greedy runs will not reproduce the headline numbers.

Strengths — where R1 Distill specifically wins

• Maths and code under your roof. At 4-bit quantisation the 14B and 32B Qwen variants are 9.0 GB and 20 GB downloads in Ollama’s library (q4_K_M), so they fit on a single 16 GB or 24 GB GPU and produce R1-style step-by-step reasoning on hardware you control.
• Permissive enough for most production use. DeepSeek’s weights are MIT; the Qwen bases are Apache 2.0 and the Llama bases carry Meta’s Llama licences. Either way, commercial use is allowed in the typical case — read the specific licence before you ship.
• Drop-in tooling. DeepSeek’s card says the distilled models “can be utilized in the same manner as Qwen or Llama models” — for instance, you can start a service with vLLM using vllm serve deepseek-ai/DeepSeek-R1-Distill-Qwen-32B --tensor-parallel-size 2 --max-model-len 32768 --enforce-eager. No custom inference stack required.
• Reasoning in a <think> block. The output contains a <think> block followed by a final answer, which is useful for debugging prompts and auditing reasoning steps. DeepSeek notes the R1 series sometimes skips it (outputting an empty <think></think>) and recommends forcing each response to start with <think>\n.

Weaknesses — where they fall short

• They are reasoning specialists, not generalists. DeepSeek’s card reports maths, science and coding benchmarks only; evaluate creative writing, multi-turn instruction following and tool use on your own tasks before relying on them.
• Quirks inherited from R1. The model card recommends setting temperature within 0.5–0.7 (0.6 is recommended) to prevent endless repetitions or incoherent outputs, avoiding system prompts and putting all instructions in the user prompt, and for maths problems including a directive such as: “Please reason step by step, and put your final answer within \boxed{}.” Ignore those settings and you will get worse output than the benchmark numbers suggest. DeepSeek also says it slightly changed the configs and tokenizers, so use the files from its repositories.
• Older base models. Qwen2.5 and Llama 3.1/3.3 are from 2024. The six distilled models do not benefit from newer base checkpoints — though DeepSeek later released DeepSeek-R1-0528-Qwen3-8B (published on Hugging Face in May 2025), which distils the chain of thought of DeepSeek-R1-0528 into Qwen3 8B Base; its model card says it surpasses Qwen3 8B by +10.0% on AIME 2024 and matches Qwen3-235B-thinking there (86.0, DeepSeek-reported). Its card also says system prompts are now supported, unlike the six originals.
• Long outputs. Reasoning traces are long: DeepSeek evaluated with a 32,768-token generation limit, and its vLLM example sets --max-model-len 32768. Budget enough context and output tokens for the thinking as well as the answer, or long traces will be cut off.

Hardware: what you actually need to run them

Weight sizes below are the 16-bit files on Hugging Face and the 4-bit (q4_K_M) files in Ollama’s library, both checked on 26 October 2026. Plan on extra memory above the file size for the KV cache, which grows with context length.

For interactive single-user inference, the 14B at Q4 on a 16–24 GB GPU is a good balance of quality and speed. The DeepSeek hardware calculator will give you a tighter estimate for your context length and batch size.

How to run R1 Distill locally

The fastest route is Ollama. After installing it, the following shell command pulls and chats with the 14B variant (DeepSeek-R1-Distill-Qwen-14B, per Ollama’s library page):

ollama run deepseek-r1:14b

Mind the tags: in Ollama’s library (26 October 2026), deepseek-r1:1.5b, 7b, 14b, 32b and 70b are the January 2025 distilled models, but deepseek-r1:8b (and latest) is DeepSeek-R1-0528-Qwen3-8B, a 5.2 GB download — not R1-Distill-Llama-8B. For the Llama-8B distill, use deepseek-r1:8b-llama-distill-q4_K_M.

For a server with concurrency, vLLM is the default. The model card ships a one-liner for the 32B:

vllm serve deepseek-ai/DeepSeek-R1-Distill-Qwen-32B \
  --tensor-parallel-size 2 \
  --max-model-len 32768 \
  --enforce-eager

SGLang works the same way:

python3 -m sglang.launch_server \
  --model deepseek-ai/DeepSeek-R1-Distill-Qwen-32B \
  --trust-remote-code --tp 2

Step-by-step instructions, including model-pull and quantisation choices, are in our running DeepSeek on Ollama walkthrough and the broader install DeepSeek locally guide.

R1 Distill vs the DeepSeek API

R1 Distill is a self-hosting story. If you want managed inference instead, DeepSeek’s API serves its current models through POST /chat/completions, the OpenAI-compatible endpoint at https://api.deepseek.com: deepseek-flash (DeepSeek-V4.1-Flash-1017) and deepseek-v4-pro (DeepSeek-V4-Pro-0813), both with open weights under MIT, a 1M-token context and up to 384K output tokens. Thinking mode is the default and returns reasoning_content alongside the final content; reasoning_effort sets low, high or max. The API is stateless — clients resend the full message history with every request. The old deepseek-reasoner ID, which first served R1 on the API, was scheduled for retirement on 2026-07-24 at 15:59 UTC and is no longer listed in the docs. See the DeepSeek API documentation for setup.

Pricing snapshot (October 2026)

R1 Distill itself is free to download and run; the only cost is your hardware. If you instead use the API, the rates as of October 2026 are below (USD per 1M tokens; off-peak is half the peak rate, and peak hours are 01:00–04:00 and 06:00–10:00 UTC, Monday to Friday, excluding Chinese public holidays). Verify on the DeepSeek API pricing page before committing budget.

Best use cases

• Maths tutoring and proof writing — see DeepSeek for math for prompt patterns that work well with reasoning models.
• Coding assistants offline — the 14B and 32B variants pair well with VS Code or a local agent; our DeepSeek for coding page covers integration patterns.
• Research workflows with private data — running locally avoids sending sensitive prompts off-machine; the DeepSeek for research guide goes deeper.
• Edge deployment — the 1.5B is a 1.1 GB download at Q4 in Ollama, small enough for low-power devices on constrained reasoning tasks; check speed on your own hardware.

Comparable alternatives

If R1 Distill is on your shortlist, you should also evaluate DeepSeek R1 vs OpenAI o1 for the API-side reasoning comparison, and the broader DeepSeek alternatives for reasoning roundup for other open reasoning checkpoints. For self-hosting flexibility specifically, the open-source AI like DeepSeek page is the right starting point. You can browse the rest of the family on the DeepSeek models hub.

Verdict

R1 Distill is the right choice when you need R1-style reasoning on hardware you control and you can live with long reasoning outputs and 2024-era base models. The 14B Qwen and 32B Qwen are the two most useful sizes for most single-GPU setups — on DeepSeek’s January 2025 numbers they beat GPT-4o-0513 and Claude-3.5-Sonnet-1022 on maths and code while fitting on one consumer or workstation GPU at Q4. For frontier-tier work, send the request to the DeepSeek API instead.