Home › Language Models › DeepSeek-V3.2-Speciale

DeepSeek-V3.2-Speciale - AI Language Models Tool

Overview

DeepSeek-V3.2-Speciale is the high‑compute, reasoning‑focused variant of the open‑weight DeepSeek‑V3.2 family. Built as a research and capability push rather than a tool‑integrated production model, Speciale emphasizes deep chain‑of‑thought reasoning and long‑context efficiency via DeepSeek Sparse Attention (DSA). The release is distributed as safetensors in BF16/FP8/F32 formats under an MIT license, with guidance to run locally using the V3.2‑Exp repository instructions. ([huggingface.co](https://huggingface.co/deepseek-ai/DeepSeek-V3.2-Speciale)) The model card and accompanying technical report describe a scaled reinforcement‑learning post‑training pipeline and a large‑scale agentic task synthesis workflow that together target improved multi‑step reasoning, compliance, and generalization for complex tasks. The authors explicitly note that Speciale is intended exclusively for deep reasoning workloads and does not support in‑model tool calling; instead, the project provides an updated chat encoding (“thinking with tools”) for environments that orchestrate external tool use outside the Speciale weights. For community verification, the release includes selected competition submissions and example assets. ([huggingface.co](https://huggingface.co/deepseek-ai/DeepSeek-V3.2-Speciale))

Model Statistics

  • Downloads: 27,281
  • Likes: 637
  • Pipeline: text-generation
  • Parameters: 685.4B

License: mit

Model Details

Architecture and scale: DeepSeek‑V3.2‑Speciale is a transformer family member in the DeepSeek V3.2 line, implemented with the project’s DeepSeek Sparse Attention (DSA) mechanism to reduce compute for long contexts; the published model card lists a ~685B parameter model and safetensors weights in BF16, F8 (F8_E4M3) and F32 formats. The model card lists the base artifact as deepseek‑ai/DeepSeek‑V3.2‑Exp‑Base and directs users to the V3.2‑Exp repo for deployment instructions. ([huggingface.co](https://huggingface.co/deepseek-ai/DeepSeek-V3.2-Speciale)) Training & alignment: The V3.2 announcement highlights a scalable reinforcement learning (post‑training) protocol and an agentic task synthesis pipeline used to generate structured multi‑step reasoning supervision at scale. The project claims this contributed to the Speciale variant’s high reasoning performance. The model card also documents a new chat encoding module (encoding/encoding_dsv32) that implements a "thinking" role encoding used during training and evaluation; note that the Speciale checkpoint itself is marked as not supporting direct tool calls. ([huggingface.co](https://huggingface.co/deepseek-ai/DeepSeek-V3.2-Speciale)) Deployment notes: DSA is intended to lower memory and compute for long‑context inference, and the community and press coverage emphasize early optimizations for non‑CUDA ecosystems (Ascend/CANN and related stacks). The Hugging Face page recommends local sampling parameters (temperature=1.0, top_p=0.95) and points to the experimental repo for hardware/inference recipes. ([tomshardware.com](https://www.tomshardware.com/tech-industry/deepseek-new-model-supports-huawei-cann))

Key Features

  • DeepSeek Sparse Attention (DSA) for reduced compute on long contexts.
  • High‑compute reasoning variant tuned with scaled RL post‑training.
  • Updated "thinking with tools" chat encoding for preserved multi‑step reasoning.
  • Weights in safetensors with BF16, FP8 (F8_E4M3), and F32 options.
  • Distributed under an MIT license; run‑locally guidance provided in V3.2‑Exp repo.

Example Usage

Example (python):

import transformers
import torch

# NOTE: DeepSeek‑V3.2‑Speciale is a very large model. Follow the project's V3.2‑Exp repo for
# production deployment and quantization instructions. The snippet below demonstrates the
# encoder/usage pattern from the model card for local experimentation only.

# 1) Install requirements: transformers, accelerate, safetensors
# 2) Clone the V3.2‑Exp repo if you need encoding/utility scripts (encoding/encoding_dsv32.py)

# Load tokenizer (uses the model's provided tokenizer)
tokenizer = transformers.AutoTokenizer.from_pretrained("deepseek-ai/DeepSeek-V3.2-Speciale")

# If you have the encoding helper from the repo (encoding/encoding_dsv32.py):
# from encoding_dsv32 import encode_messages, parse_message_from_completion_text

messages = [
    {"role": "user", "content": "Prove that every prime of the form 4k+1 is a sum of two squares."}
]

# Example: use encode_messages if you follow the repo's encoding conventions
# encode_config = dict(thinking_mode="thinking", drop_thinking=True, add_default_bos_token=True)
# prompt = encode_messages(messages, **encode_config)

# Fallback: simple prompt string (the model card provides repository encoders for best results)
prompt = "User: " + messages[0]["content"] + "\nAssistant: "

# Tokenize
inputs = tokenizer(prompt, return_tensors="pt")

# Load model (toy example: device_map='auto' and bfloat16 dtype where supported)
model = transformers.AutoModelForCausalLM.from_pretrained(
    "deepseek-ai/DeepSeek-V3.2-Speciale",
    torch_dtype=torch.bfloat16,
    device_map='auto',
    trust_remote_code=True
)

# Generate with the sampling parameters recommended in the model card
generation_config = dict(max_new_tokens=512, temperature=1.0, top_p=0.95)
outputs = model.generate(**inputs, **generation_config)

print(tokenizer.decode(outputs[0], skip_special_tokens=True))

# IMPORTANT: This example is for illustration. Refer to the DeepSeek V3.2‑Exp repository for
# precise encoding/parsing helpers, quantization recipes, and hardware‑specific deployment steps.
# See the model card and the V3.2‑Exp repo for details. ([huggingface.co](https://huggingface.co/deepseek-ai/DeepSeek-V3.2-Speciale))

Benchmarks

Parameters: ≈685 billion (Source: https://huggingface.co/deepseek-ai/DeepSeek-V3.2-Speciale)

Downloads (last month, as shown on Hugging Face): 27,281 (Source: https://huggingface.co/deepseek-ai/DeepSeek-V3.2-Speciale)

Community likes (Hugging Face): 637 likes (Source: https://huggingface.co/deepseek-ai/DeepSeek-V3.2-Speciale)

Competition / reasoning claims (per model card): Model card reports gold‑medal performance on IMO 2025 and IOI 2025; claims to surpass GPT‑5 on reasoning benchmarks (Source: https://huggingface.co/deepseek-ai/DeepSeek-V3.2-Speciale)

Press coverage (long‑context / DSA): Independent reports highlight DSA, long‑context efficiency, and support for non‑CUDA stacks (Source: https://www.reuters.com/technology/deepseek-releases-model-it-calls-intermediate-step-towards-next-generation-2025-09-29/)

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.