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

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

Overview

DeepSeek-V3.2-Exp is an experimental, open-weight large language model from DeepSeek that introduces DeepSeek Sparse Attention (DSA) to reduce compute and memory for long-context training and inference while preserving output quality comparable to V3.1-Terminus. The release is provided under an MIT license with safetensor weights available in multiple tensor formats (BF16, FP8/E4M3, F32) and is published on Hugging Face with an updated inference demo and conversion tools for local deployment. (Source: Hugging Face model card: https://huggingface.co/deepseek-ai/DeepSeek-V3.2-Exp) Designed as an intermediate step toward DeepSeek's next-generation models, V3.2-Exp keeps training configurations aligned with V3.1 to allow apples-to-apples comparison; public benchmark tables published by DeepSeek show near-parity with V3.1 across reasoning and agentic tool-use tasks. Operationally, V3.2-Exp has day‑0 support on ecosystem runtimes such as SGLang and vLLM and ships example scripts and Docker images to run on H200/MI350/various NPUs. DeepSeek also published open-source GPU kernels (TileLang, DeepGEMM, FlashMLA) and a tech report for researchers. (Sources: Hugging Face model card; DeepSeek API docs; DeepSeek GitHub.)

Model Statistics

Downloads: 69,742
Likes: 932
Pipeline: text-generation
Parameters: 685.4B

License: mit

Model Details

Architecture and size: DeepSeek-V3.2-Exp is a ~685B-parameter transformer-class model (checkpoint listed as 685B on the Hugging Face model card) built from the same structural family as V3.1-Terminus, with a new sparse attention component (DeepSeek Sparse Attention, DSA) layered into the attention/kv/indexer pipeline to prune unnecessary token-to-token computation in long windows. (Source: Hugging Face: https://huggingface.co/deepseek-ai/DeepSeek-V3.2-Exp) Sparsity mechanism & training: DSA uses a trainable indexer that learns to approximate dense head-summed attention distributions (indexer warm-up followed by sparse training using KL losses and large-scale token counts, per DeepSeek technical notes and third-party writeups). The implementation is designed for MLA/MQA decoding modes and KV reuse to maximize throughput. GPU/kernel support includes TileLang (research-friendly kernels), DeepGEMM (indexer logits and paged kernels), and FlashMLA (sparse attention kernels). (Sources: DeepSeek GitHub tech report; technical coverage: https://github.com/deepseek-ai/DeepSeek-V3.2-Exp and analysis articles.) Formats & deployment: The authors provide safetensor checkpoints in multiple numeric formats (BF16, FP8/E4M3, F32). Deployments are supported with example conversion and torchrun scripts, SGLang Docker images (H200/MI350/NPUs), and day‑0 vLLM recipes. The model card also records community adoption signals (downloads and likes) and a recent inference-code RoPE fix published on 2025-11-17. (Sources: Hugging Face; SGLang docs; repository updates.)

Key Features

DeepSeek Sparse Attention (DSA) — trainable fine-grained sparse attention for long contexts.
Open-weight safetensors in BF16, FP8 (E4M3), and F32 formats for flexible precision.
Day‑0 support in vLLM and SGLang with Docker images and example launch commands.
Conversion and torchrun inference demo for multi-GPU model-parallel local hosting.
Open-source performance kernels: TileLang, DeepGEMM, and FlashMLA for research and production.

Example Usage

Example (python):

import transformers

# Example: encode DeepSeek-style chat messages using DeepSeek's encoding helper
# (adapted from DeepSeek V3.2 examples; use encoding_dsv32 utilities from the repo)

from encoding_dsv32 import encode_messages, parse_message_from_completion_text

# load tokenizer from Hugging Face (tokens/encoding compatible with DeepSeek tooling)
tokenizer = transformers.AutoTokenizer.from_pretrained("deepseek-ai/DeepSeek-V3.2")

messages = [
    {"role": "user", "content": "Summarize the main benefits of sparse attention."},
]

encode_config = dict(thinking_mode="thinking", drop_thinking=True, add_default_bos_token=True)
prompt = encode_messages(messages, **encode_config)

# Tokenize and inspect
tokens = tokenizer.encode(prompt)
print("Prompt tokens:", tokens)

# Note: DeepSeek provides inference examples and a conversion script in the `inference/` folder
# for running the model locally (torchrun + generate.py), and also documents SGLang/vLLM recipes.
# See: https://huggingface.co/deepseek-ai/DeepSeek-V3.2-Exp and the model GitHub for full runbook.

Pricing

DeepSeek announced a >50% API price cut for its hosted API at the V3.2-Exp launch; exact commercial pricing tiers are not published in the model card — see DeepSeek API docs for details. (Source: DeepSeek API announcement: https://api-docs.deepseek.com/news/news250929; coverage: Reuters.)