Home › Language Models › DeepSeek V3.1

DeepSeek V3.1 - AI Language Models Tool

Overview

DeepSeek V3.1 is an open-source, agent-oriented large language model released by DeepSeek-AI and published on Hugging Face under an MIT license. It is a Mixture-of-Experts (MoE) decoder-only transformer family member that exposes both a base checkpoint (DeepSeek-V3.1-Base) and a post-trained chat/tuned variant. The release emphasizes agent readiness: built-in chat templates for “thinking” and “non-thinking” modes, explicit tool-calling formats, and example code/search agent trajectories to simplify building multi-step tool-using systems (search, code agents, and function-calling workflows). According to the model card, V3.1 was further pre-trained with large-scale long-context extension phases to support a 128K token context window, and the team optimized the model for improved tool use and faster “thinking” mode responses (Hugging Face model card; DeepSeek technical report). Technically, DeepSeek V3.1 targets developers who need long-context reasoning, multi-step agents, or affordable self-hosting. The project publishes both full weights and quantized variants (BF16, F32, and FP8/E4M3 formats) and provides guidance for FP8/UE8M0 usage. Community benchmarks on the model card show strong performance on coding benchmarks and multi-step agent tasks in the authors’ evaluation suite, while independent reporting and community threads also document trade-offs (accuracy and safety audits, regulatory scrutiny, and mixed user feedback). The model remains fully downloadable for on-premise use and is integrated into DeepSeek’s API offering for managed inference.

Model Statistics

  • Downloads: 52,635
  • Likes: 811
  • Pipeline: text-generation
  • Parameters: 684.5B

License: mit

Model Details

Architecture and scale: DeepSeek V3.1 is a 671B-parameter Mixture-of-Experts (MoE) decoder-only transformer with roughly 37B activated parameters per token (the MoE routing activates a small subset of experts at inference). The V3 family uses DeepSeekMoE and Multi-head Latent Attention (MLA) techniques and was trained under a multi-token prediction objective (DeepSeek V3 technical report). Pretraining and long-context extension: the V3 technical report and model card state the family was pre-trained on very large corpora (DeepSeek-V3 report cites ~14.8 trillion tokens for V3) and V3.1 adds two long-context extension phases — a 32K phase expanded to ~630B tokens and a 128K phase expanded to ~209B tokens — enabling a 128K token context window for both base and chat variants (Hugging Face model card; arXiv technical report). Precision, deployment, and tooling: weights are published in BF16 and multiple quantized formats (F8_E4M3 / UE8M0-compatible FP8 formats), and the maintainers document FP8 handling recommendations (e.g., keep some gate/MLP parameters in FP32). The model card includes templated chat formats (non-thinking vs thinking tokens), a strict tool-call schema for function/agent integration, and example code-agent/search-agent trajectories. Training and evaluation: the V3 family reports internal benchmark comparisons (MMLU variants, code benchmarks, SWE-bench, LiveCodeBench, math contest pass@1 scores) showing improved agent and code performance versus earlier DeepSeek checkpoints. The model is released under MIT license and available via Hugging Face and ModelScope; DeepSeek also offers managed API endpoints (chat / reasoner) for hosted access.

Key Features

  • 128K token context window for single-call long-document reasoning and retrieval.
  • Hybrid 'thinking' and 'non-thinking' chat templates to toggle chain-of-thought behavior.
  • Mixture-of-Experts (MoE) 671B total / ~37B activated parameters for cost-effective inference.
  • Built-in tool-call schema and example code/search agent templates for multi-step agents.
  • Published weights in BF16/F32 and FP8 (UE8M0) formats with quantization guidance.
  • Open-source MIT license; weights downloadable from Hugging Face and ModelScope.

Example Usage

Example (python):

from transformers import AutoTokenizer, AutoModelForCausalLM
import torch

# Example: load tokenizer and model (requires sufficient hardware or quantized weights)
MODEL = "deepseek-ai/DeepSeek-V3.1"

tokenizer = AutoTokenizer.from_pretrained(MODEL)
model = AutoModelForCausalLM.from_pretrained(MODEL, torch_dtype=torch.bfloat16, device_map='auto')

# Build chat messages and apply DeepSeek chat template (thinking=True for chain-of-thought)
messages = [
    {"role": "system", "content": "You are a helpful assistant."},
    {"role": "user", "content": "Summarize key points from the pasted document and list follow-ups."},
]
# The tokenizer provides an apply_chat_template helper on the model card
prompt = tokenizer.apply_chat_template(messages, tokenize=False, thinking=True, add_generation_prompt=True)
inputs = tokenizer(prompt, return_tensors='pt')

# Generate (small example; adjust generation parameters for real workloads)
out = model.generate(**{k: v.to(model.device) for k, v in inputs.items()}, max_new_tokens=256, do_sample=False)
print(tokenizer.decode(out[0], skip_special_tokens=True))

# Note: This example assumes local hardware with enough memory or quantized weights. See the model card for FP8/UE8M0 guidance and agent/tool-call formats.

Pricing

DeepSeek publishes both downloadable open-source weights and managed API endpoints. DeepSeek’s official API docs list metered pay-as-you-go rates (separate input/output token billing) for their managed endpoints (deepseek-chat / deepseek-reasoner) and show low per‑1M‑token rates compared with many competitors; the API docs give explicit per‑million-token prices and cache-hit / cache-miss distinctions (see DeepSeek API docs). DeepSeek has also offered off‑peak discounts in the past and published pricing pages for models and modes. (Sources: DeepSeek API documentation; DeepSeek Hugging Face model card; industry reporting on off‑peak pricing.)

Benchmarks

MMLU-Redux (Exact Match) — DeepSeek V3.1 (Non-Thinking): 91.8 (Source: https://huggingface.co/deepseek-ai/DeepSeek-V3.1)

LiveCodeBench (Pass@1) — DeepSeek V3.1 (Non-Thinking): 56.4% (Source: https://huggingface.co/deepseek-ai/DeepSeek-V3.1)

LiveCodeBench (Pass@1) — DeepSeek V3.1 (Thinking): 74.8% (Source: https://huggingface.co/deepseek-ai/DeepSeek-V3.1)

AIME 2024 (Pass@1): 66.3% (Source: https://huggingface.co/deepseek-ai/DeepSeek-V3.1)

SWE Verified (Agent mode): 66.0% (Source: https://huggingface.co/deepseek-ai/DeepSeek-V3.1)

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.