Home › Language Models › DeepSeek V3.1

DeepSeek V3.1 - AI Language Models Tool

Overview

DeepSeek V3.1 is an open-source, long-context large language model optimized for chat, tool use, and agent-style workflows. It is released under the MIT license and provided in both a Base checkpoint and a post‑trained chat model; the project emphasizes a hybrid “thinking” and “non-thinking” chat template (switchable by prompt) and explicit tool‑call formatting for deterministic agent integration. The model supports an unusually large 128,000 token context window, enabling single-pass reasoning over very long documents, full code repositories, or multi‑document research contexts. ([huggingface.co](https://huggingface.co/deepseek-ai/DeepSeek-V3.1)) Technically the released V3.1 family is a Mixture‑of‑Experts (MoE) design with 671B total parameters and about 37B activated parameters per token; the distribution and training strategy emphasize long‑context extension and efficient inference (FP8/UE8M0 compatibility). The V3.1 release notes and model card call out improvements in tool calling and faster “thinking” mode responses versus earlier checkpoints, while community feedback (forum and Reddit threads) has been mixed — strong praise for code and long‑context tasks but reports of style/regeneration regressions in some conversational/creative settings. ([huggingface.co](https://huggingface.co/deepseek-ai/DeepSeek-V3.1))

Model Statistics

  • Downloads: 149,522
  • Likes: 818
  • Pipeline: text-generation

License: mit

Model Details

Architecture and scale: DeepSeek V3.1 is presented as a Mixture‑of‑Experts (MoE) transformer family with 671 billion total parameters and ~37 billion activated parameters per token, matching the architecture described in the DeepSeek V3 technical report. The model family includes a V3.1-Base checkpoint and a post‑trained V3.1 conversational checkpoint. ([arxiv.org](https://arxiv.org/abs/2412.19437)) Context extension and training: V3.1 uses a two‑phase long‑context extension: a 32K extension phase scaled to ~630B tokens and a 128K extension phase expanded to ~209B tokens (combined additional pretraining on the order of ~840B tokens beyond the base), enabling a 128K token usable context window. The authors report pretraining on very large corpora (the V3 technical report cites ~14.8 trillion tokens for the V3 training lineage) and post‑training steps (SFT/RL) for chat and tool behaviors. ([huggingface.co](https://huggingface.co/deepseek-ai/DeepSeek-V3.1)) Precision & inference: V3.1 weight and activation checkpoints are provided in FP8 UE8M0 format for compatibility with microscaling inference stacks; model guidance recommends loading some gate/score corrections in FP32 for numerical stability. The Hugging Face model card includes chat templates (Non‑Thinking, Thinking, ToolCall) and a strict tool‑call format for deterministic function calling and agent chaining. ([huggingface.co](https://huggingface.co/deepseek-ai/DeepSeek-V3.1)) Agent/tool support: The model card supplies code‑agent and search‑agent examples and prescribes a precise tool‑call token framing (special tool call begin/end tokens and JSON argument schema) to enable multi‑turn tool use and search workflows within the 128K context window. ([huggingface.co](https://huggingface.co/deepseek-ai/DeepSeek-V3.1))

Key Features

  • 128,000 token context window for single-pass long-document reasoning and codebase analysis.
  • Mixture‑of‑Experts architecture: 671B total parameters, ~37B activated per token (cost-efficient MoE).
  • Hybrid chat templates: switchable Thinking (chain-of-thought style) and Non‑Thinking modes.
  • Explicit tool‑call format for deterministic function calling and multi-turn agent workflows.
  • FP8 (UE8M0) formatted weights and activations for microscaling / efficient inference.
  • Post‑trained conversational checkpoint plus a Base checkpoint for custom fine‑tuning.
  • Released under MIT license; model weights and assets available on Hugging Face.

Example Usage

Example (python):

from transformers import AutoTokenizer, AutoModelForCausalLM, TextGenerationPipeline

# Example follows the usage pattern shown in the model card: use the tokenizer's chat template helpers
# and generate in 'thinking' or 'non-thinking' mode. Adjust device_map/precision for your hardware.

tokenizer = AutoTokenizer.from_pretrained("deepseek-ai/DeepSeek-V3.1")
model = AutoModelForCausalLM.from_pretrained(
    "deepseek-ai/DeepSeek-V3.1",
    trust_remote_code=True,          # required for some vendor-provided model code
    torch_dtype="auto",
)

# Build a message sequence (system + user) and apply the provided chat template helpers
messages = [
    {"role": "system", "content": "You are a helpful assistant."},
    {"role": "user", "content": "Summarize the following 40-page contract: <DOCUMENT_HERE>"},
]

# Many DeepSeek distributions include a tokenizer helper to format messages for Thinking mode
try:
    formatted = tokenizer.apply_chat_template(messages, tokenize=False, thinking=True, add_generation_prompt=True)
except Exception:
    # Fallback: join as plain text if helper isn't available in your transformers version
    formatted = "\n".join([f"{m['role']}: {m['content']}" for m in messages])

# Generate (simple example). For large-context runs, ensure you use an engine that supports 128K.
pipeline = TextGenerationPipeline(model=model, tokenizer=tokenizer, device=0)
outputs = pipeline(formatted, max_new_tokens=512, do_sample=False)
print(outputs[0]['generated_text'])

# Note: See the official model card for the precise token framing, tool-call syntax, and
# recommended precision/FP8 handling when running large-context loads. Hugging Face model card documents these details.

Benchmarks

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

AIME 2025 (Pass@1): 49.8 (Source: https://huggingface.co/deepseek-ai/DeepSeek-V3.1)

HMMT 2025 (Pass@1): 33.5 (Source: https://huggingface.co/deepseek-ai/DeepSeek-V3.1)

Total params / Activated params: 671B total, 37B activated (Source: https://arxiv.org/abs/2412.19437)

Downloads (last month, as shown on model card): 123,230 (last month, model card snapshot) (Source: https://huggingface.co/deepseek-ai/DeepSeek-V3.1)

Last Refreshed: 2026-03-03

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.