Home › Language Models › DeepSeek-R1-Distill-Llama-8B

DeepSeek-R1-Distill-Llama-8B - AI Language Models Tool

Overview

DeepSeek-R1-Distill-Llama-8B is an 8-billion-parameter distilled language model in DeepSeek’s R1 family, produced by fine-tuning Llama-3.1-8B with reasoning samples generated by the larger DeepSeek-R1. It is designed to provide strong chain-of-thought reasoning, math, and code performance in a compact, self-hostable checkpoint suitable for research and production use. According to the official model card, the distillation pipeline uses DeepSeek-R1 outputs and selective tokenizer/config adjustments to transfer reasoning behaviours into smaller dense checkpoints. ([huggingface.co](https://huggingface.co/deepseek-ai/DeepSeek-R1-Distill-Llama-8B)) The model is distributed under an MIT license and is available on the Hugging Face Hub as a safetensors checkpoint (BF16 weights) with a 32k-token context window, making it appropriate for long-form reasoning and document-level tasks when paired with an inference engine that supports large contexts. The project’s public evaluation table reports competitive pass@1 and reasoning-centered results for the distilled Llama-8B variant on benchmarks such as AIME and MATH-500, showing the practical value of the RL-driven R1 pipeline for creating smaller but capable models. ([huggingface.co](https://huggingface.co/deepseek-ai/DeepSeek-R1-Distill-Llama-8B))

Model Statistics

  • Downloads: 574,861
  • Likes: 835
  • Pipeline: text-generation
  • Parameters: 8.0B

License: mit

Model Details

Architecture and provenance: DeepSeek-R1-Distill-Llama-8B is a distilled derivative of Llama-3.1-8B (base) that was fine-tuned using data generated by DeepSeek-R1’s reinforcement-learning-driven pipeline; the model weights are released under the MIT license. The Hugging Face model card lists the checkpoint as an 8B-parameter BF16 safetensors file and explicitly notes that distilled variants were produced by using R1-generated samples and slight tokenizer/config changes. ([huggingface.co](https://huggingface.co/deepseek-ai/DeepSeek-R1-Distill-Llama-8B)) Context and inference: The model supports up to 32,768 tokens (32k) for generation and was benchmarked with recommended sampling parameters of temperature ≈0.6 and top-p ≈0.95. The maintainers recommend avoiding an external system prompt and placing instructions in the user prompt for best behaviour; they also document a specific "thinking" marker (<think>) to encourage chain-of-thought outputs when needed. ([huggingface.co](https://huggingface.co/deepseek-ai/DeepSeek-R1-Distill-Llama-8B)) Compatibility and deployment: The model is intended to run similarly to other Llama/Qwen checkpoints and is commonly deployed with inference engines such as vLLM or SGLang (examples and CLI invocation are provided in the model card). Quantized/community builds (GGUF/4-bit) and many finetunes/adapter variants exist in the Hugging Face model tree and Spaces ecosystem. ([huggingface.co](https://huggingface.co/deepseek-ai/DeepSeek-R1-Distill-Llama-8B))

Key Features

  • 8B-parameter distilled Llama-3.1 checkpoint optimized with R1-generated reasoning data.
  • 32k-token context window for long-form reasoning and document-level tasks.
  • BF16 safetensors weights suitable for high-throughput GPU inference.
  • Tuned for chain-of-thought / step-by-step reasoning via RL-derived samples.
  • Benchmarked on math and code suites (AIME, MATH-500, LiveCodeBench) with public scores.
  • Provided under MIT license and compatible with vLLM, SGLang, and many community quantizations.

Example Usage

Example (python):

from vllm import LLM, SamplingParams

# Example: local inference with vLLM (downloads model from Hugging Face Hub).
# Install: pip install vllm

sampling_params = SamplingParams(temperature=0.6, top_p=0.95, max_tokens=512)
llm = LLM(model="deepseek-ai/DeepSeek-R1-Distill-Llama-8B", dtype="auto", trust_remote_code=True)

prompt = "Please reason step-by-step and solve: If 3x+5=20, what's x?"
outputs = llm.generate([prompt], sampling_params)

for out in outputs:
    # vLLM returns a list of RequestOutput objects; text lives at outputs[0].outputs[0].text
    print(out.outputs[0].text)

# Note: DeepSeek R1 distill models are often served via vLLM or SGLang; see the model card for
# recommended runtime flags (max_model_len, tensor-parallel, enforce-eager, etc.).

Pricing

Model weights are published under the MIT license and are free to download and self-host. DeepSeek also offers a pay-as-you-go API with published token pricing (example: input cache-hit $0.028 / 1M tokens, input cache-miss $0.28 / 1M tokens, output $0.42 / 1M tokens) — see DeepSeek API docs for current rates and enterprise options. Self-hosting infrastructure costs vary by hardware and cloud provider. ([huggingface.co](https://huggingface.co/deepseek-ai/DeepSeek-R1-Distill-Llama-8B))

Benchmarks

AIME 2024 (pass@1): 50.4 (Source: https://huggingface.co/deepseek-ai/DeepSeek-R1-Distill-Llama-8B (model card evaluation table). ([huggingface.co](https://huggingface.co/deepseek-ai/DeepSeek-R1-Distill-Llama-8B)))

AIME 2024 (cons@64): 80.0 (Source: https://huggingface.co/deepseek-ai/DeepSeek-R1-Distill-Llama-8B (distilled model evaluation). ([huggingface.co](https://huggingface.co/deepseek-ai/DeepSeek-R1-Distill-Llama-8B)))

MATH-500 (pass@1): 89.1 (Source: https://huggingface.co/deepseek-ai/DeepSeek-R1-Distill-Llama-8B (distilled model evaluation). ([huggingface.co](https://huggingface.co/deepseek-ai/DeepSeek-R1-Distill-Llama-8B)))

LiveCodeBench (pass@1-COT): 39.6 (Source: https://huggingface.co/deepseek-ai/DeepSeek-R1-Distill-Llama-8B (distilled model evaluation). ([huggingface.co](https://huggingface.co/deepseek-ai/DeepSeek-R1-Distill-Llama-8B)))

Codeforces (rating, distilled Llama-8B): 1205 (Source: https://huggingface.co/deepseek-ai/DeepSeek-R1-Distill-Llama-8B (distilled model evaluation). ([huggingface.co](https://huggingface.co/deepseek-ai/DeepSeek-R1-Distill-Llama-8B)))

Hugging Face - Downloads last month: 574,861 (Source: Model page statistics on Hugging Face. ([huggingface.co](https://huggingface.co/deepseek-ai/DeepSeek-R1-Distill-Llama-8B)))

Hugging Face - Likes: 835 (Source: Model page metadata on Hugging Face. ([huggingface.co](https://huggingface.co/deepseek-ai/DeepSeek-R1-Distill-Llama-8B)))

Last Refreshed: 2026-01-09

Key Information

  • Category: Language Models
  • Type: AI Language Models Tool
Visit Official Website

Related Tools

Qwen2.5-7B

Qwen2.5-7B is a large language model designed for text generation, featuring improvements in coding, mathematics, instruction following, long text generation, and multilingual support. It supports context lengths up to 128K tokens and is intended for sophisticated NLP tasks.
DeepSeek-V3

Open‑weight flagship LLM (base and chat) with strong reasoning and tool‑use; runs with SGLang/LMDeploy/vLLM and permissive model license.
Llama 3

Llama 3 is an open access large language model (LLM) released by Meta, available in various configurations (8B and 70B parameters) with capabilities for fine-tuning and integrations into platforms like Hugging Face, Google Cloud, and Amazon SageMaker.
UNfilteredAI-1B

A large-scale text generation model designed for creative and unconstrained content production without traditional filtering.
Shuttle-3

Shuttle-3 is a state-of-the-art language model designed for high-quality text generation, particularly suited for complex chat, multilingual communication, and reasoning tasks. It is fine-tuned from the Qwen-2.5-72b-Instruct model and designed to emulate high-quality prose similar to Claude 3 models.
WizardLM

WizardLM is a state-of-the-art large language model designed for complex chat, multilingual tasks, reasoning, and agent functionalities. It features an AI-powered pipeline (Auto Evol-Instruct) that optimizes instruction datasets for improved performance across various domains and leverages Arena Learning for an expanded learning pool of challenging instruction data.