DeepSeek-V3.2-Exp has arrived as an experimental iteration that focuses on efficiency at scale and in long contexts, with a clear nod to the community: open source, published kernels, and a cheaper API.
Beyond the noise, the key is that this model, based on V3.1-Terminus, debuts a fine-grained sparse attention mechanism called DeepSeek Sparse Attention (DSA) that accelerates training and inference while maintaining quality. The company has already launched it in App, Web and API, and has cut usage prices by more than 50 %, an aggressive move that, frankly, squeezes the competition.
Key new features of DeepSeek-V3.2-Exp
The star innovation is DSA, which allows a selective attention on relevant parts of the context without exhaustively going through the entire sequence. According to the company itself, the impact on quality is very low, while the efficiency gain in long-context it is tangible.
In availability, the model is operational in application, web and API from day one, accompanied by a significant price drop (50%+) to facilitate testing and adoption. For those who want to compare, DeepSeek maintains a temporary endpoint from V3.1-Terminus until October 15, 2025 at 15:59 UTC.
DeepSeek-V3.2-Exp performance and benchmarks: parity with V3.1-Terminus
DeepSeek has searched for the parity with V3.1-Terminus on a wide range of tests, precisely to isolate the effect of introducing scattered attention. In practice, this results in comparable metrics in reasoning, coding and use of agent-type tools.
Various sources provide figures that help to set expectations: V3.2-Exp is described as a model with 685 billion parameters and similar performance or with slight variations depending on the domain. In reasoning without tools, numbers such as 85.0 are cited MMLU-Pro There 89.3 in AIME 2025; in agent scenarios, 40.1 appear in BrowseComp There 67.8 in SWE VerifiedThese are results that fit with the official narrative of validate efficiency rather than pursuing a gross leap in accuracy.
There are even fine comparisons: in coding tasks, a rise to 2121 in Codeforces compared to 2046, while in more humanistic tests slight drops are observed (e.g., 19.8 compared to 21.7 in Humanity's Last Exam). Overall, the table suggests equilibrium: specific improvements and small concessions, with speed as the main focus.
DSA: Fine-grained scattered attention, told clearly
Classical care becomes expensive with broad contexts; DSA Minimizes work where it contributes little. By applying sparsity with fine-grained control, the model concentrates computation where it actually finds signal, improving the latency and reducing consumption, without distorting the output.
At the level of real experience, this is noticeable in tasks that require a lot of context: long document summaries, log analysis, agents holding long dialogues, or pipelines that mix retrieval and generation. Right there, the efficiency It's not a luxury: it's the difference between something being usable at scale or not.
DeepSeek-V3.2-Exp availability, pricing, and comparisons
DeepSeek has announced that V3.2-Exp is now available on App, Web and API. In addition, it has cut the price of the API by more than 50 % with immediate effect, a decision that aims to expand adoption and encourage comparative tests.
For those who want to compare with the previous model, V3.1-Terminus is maintained in a end point temporary until 10/15/2025 15:59 (UTC). The company also invites submissions feedback through a public form, reinforcing a dynamic of continuous improvement with the community.
Open source status: weights, technical report, and kernels
DeepSeek publishes the model in Hugging Face, along with a technical report that documents changes and results. There is a clear commitment to transparency and to promoting long-term applied research with lower costs.
At the kernel level, there are two ways: TileLang for reading and prototyping more accessible, and CUDA for maximum performance. Logit index kernels (including paged variants) are in DeepGEMM, while those with scattered attention are published in FlashMLAThis separation makes it easier for the research and production-oriented communities to find their fit.
Local execution of DeepSeek-V3.2-Exp and inference demos
DeepSeek offers a repository of inference with an updated demo to get started quickly and inspect the architecture. The first step is to convert the Hugging Face weights to the format expected by the demo, defining the number of experts and model parallelism.
Example commands for interactive conversion and generation (set EXPERTS=256 and MP to the number of GPUs): can be used as is in a prepared environment.
cd inference export EXPERTS=256 python convert.py --hf-ckpt-path ${HF_CKPT_PATH} --save-path ${SAVE_PATH} --n-experts ${EXPERTS} --model-parallel ${MP} export CONFIG=config_671B_v3.2.json torchrun --nproc-per-node ${MP} generate.py --ckpt-path ${SAVE_PATH} --config ${CONFIG} --interactive
For those who prefer SGLang, there are prepared images and a boot command. Support covers NVIDIA (H200), AMD (MI350) GPUs, and certain NPUs, with specific tags.
# H200 docker pull lmsysorg/sglang:dsv32 # MI350 docker pull lmsysorg/sglang:dsv32-rocm # NPUs docker pull lmsysorg/sglang:dsv32-a2 docker pull lmsysorg/sglang:dsv32-a3 python -m sglang.launch_server --model deepseek-ai/DeepSeek-V3.2-Exp --tp 8 --dp 8 --page-size 64
If you prefer vLLM, has support from day one. It is advisable to review the official recipes for up-to-date parameters and optimizations by hardware.
API: Endpoints, Compatibility, and Expirations
The API of DeepSeek follows standard conventions and is compatible with popular SDKs. By default, using the base URL https://api.deepseek.com you are targeting V3.2-Exp, which simplifies initial integration and access to the reduced rate.
For benchmarking, there is a temporary endpoint for V3.1-Terminus: https://api.deepseek.com/v3.1_terminus_expires_on_20251015. Keep in mind the expiration date and time (October 15, 2025, 15:59 UTC) to plan benchmarks.
In addition, there is compatibility with the ecosystem of anthropic. You can use the base https://api.deepseek.com/anthropic for Claude-style interactions, or the variant associated with the temporary endpoint if you need to compare with the previous model.
Authentication and key management
Requests are authenticated by bearer in the Authorization header. Generate your key from the DeepSeek dashboard and store it securely, for example, in environment variables or file managers. secrets as AWS Secrets Manager.
The panel shows usage and billing to control consumption of token usage. Although prices have dropped, it is advisable to apply rate limiting and periodic key rotation on computers, in addition to revoking any compromised key Without delay.
Chat completions, templates, and basic requests
The central endpoint is /chat/completions, which processes multi-turn dialogues and maintains context between calls, an ideal scenario for V3.2-Exp's long-context strengths. There are two typical model modes: deepseek-chat y deepseek-reasoner.
A simple request body might look like this, using escaped JSON (represented here as " for clarity): includes a system prompt and a user prompt.
{ "model": "deepseek-chat", "messages": [ { "role": "system", "content": "You are a technical expert." }, { "role": "user", "content": "Explain sparse attention." } ], "stream": false }
When you want real-time answers, activate stream=true. The headers must include Content-Type: application/json and the Authorization token: Bearer ${DEEPSEEK_API_KEY}. If you're working with explicit reasoning, you can control the behavior with the flag reasoning.enabled.
Response structure and SSE streaming
Non-streaming responses include fields such as id, object, created, model, choices, and usage. In choices, you'll find the generated content (role: "assistant"), and in usage, the details of prompt_tokens, completion_tokens and total_tokens.
In streaming mode, the API sends Server-Sent Events. Each fragment arrives as a data event with a delta that you must accumulate. This is the ideal option for interfaces interactive or terminals with incremental output.
Function calling and output in strict JSON
You can define tools so the model decides when to call a function, for example, to retrieve data or execute actions. This fits well with agent flows and integrations. backend.
If you need structured output, force JSON mode using response_format. This is useful for data extraction or validation automatic in pipelines.
Python examples with OpenAI-style SDK
With Python, the input curve is very smooth. Set the api_base like DeepSeek, define the key and launch requests; you can switch between standard and streaming mode depending on your use case.
import openai openai.api_base = "https://api.deepseek.com" openai.api_key = "your_api_key_here" response = openai.ChatCompletion.create( model="deepseek-chat", messages=[ {"role": "system", "content": "You are a coding assistant."}, {"role": "user", "content": "Write a Python function to calculate Fibonacci numbers."} ], stream=False ) print(response.choices[0].message.content) # Streaming stream = openai.ChatCompletion.create(model="deepseek-chat", messages=[...], stream=True) for chunk in stream: if chunk.choices[0].delta.content is not None: print(chunk.choices[0].delta.content, end="") # Function calling (tool definition) tools = [ { "type": "function", "function": { "name": "get_weather", "description": "Get current weather", "parameters": { "type": "object", "properties": { "location": {"type": "string"} }, "required": [ "location" ] } } } ]
For strict JSON output, set response_format a {"type": "json_object"}. And if you're going to have long dialogues, remember to go narrowing down the context to keep you within limits and maximize efficiency.
Integration with Apidog
Apidog accelerates the prototyping Callback: Import specifications, save environment variables (such as the key), build POSTs, and test on the fly. Its response simulator makes it easy to test extreme cases at no additional cost. token usage.
It also generates code snippets in different idioms and offers a timeline view for debugging authentication or parameters. Since V3.2-Exp handles broad contexts, Apidog is a great way to experiment with long prompts and see performance.
Good practices to get the most out of it
Define system prompts clear and concise that delimit behavior. For complex problems, the reasoning mode can help, combining it with thought-structuring techniques appropriate to your case.
Manage the context with head: Although V3.2-Exp tolerates long-context (cases up to 128K are cited), excessive history can penalize efficiency. It implements smart truncation, hidden for frequent and batch queries where it makes sense.
In security, it sanitizes input to prevent prompt injections and logs interactions to audit. Adjust temperature and top_p according to your goal: low values ​​for determinism, high values ​​for creativity.
Perform A/B tests between deepseek-chat y deepseek-reasoner to choose the optimal mode. And remember the rate limit to avoid surprises in the billing.
Comparison with V3.1-Terminus
The introduction of DSA brings improvements in latency which, in some scenarios, approach 3x speed without sacrificing overall quality parity. It's an evolution focused on the relationship power/efficiency more than in accuracy records.
The small increases in coding and slight decreases in humanities domains reflect the fine-tuning of a model that, by design, is experimentalThe V3.1 temporal endpoint allows for direct comparisons showing DSA gains in the long run.
Advanced local deployment
For privacy-sensitive or offline deployments, download the pesos from Hugging Face and using the official conversion scripts is the way to go. Set the number of experts (e.g. 256) and adjust the model parallelism to your needs. GPUs.
The inference demo allows testing in interactive mode, and the kernels in TileLang or CUDA will help you squeeze performance according to priorities: speed of prototyping or maximum throughput in production.
Open kernels and performance
TileLang prioritizes readability and design for research, so you can quickly iterate on new ideas. It's perfect if you're exploring care variants. scattered or memory optimizations.
To squeeze every millisecond, CUDA kernels come into play: logit indices (with paged versions) are in DeepGEMM, while those with scattered attention live in FlashMLAThis segmentation allows each team to choose the stack optimal without redoing work.
DeepSeek-V3.2-Exp License, Appointment and Contact
The repository and the pesos The model is published under the MIT license. This opens the door to commercial uses with great flexibility, encouraging adoption and innovation in the ecosystem.
To reference V3.2-Exp in jobs, DeepSeek provides a appointment entry of type @misc with the title «DeepSeek-V3.2-Exp: Boosting Long-Context Efficiency with DeepSeek Sparse Attention» and authorship «DeepSeek-AI» (year 2025). For questions or incidents, the contact email is service@deepseek.com.
Official resources and useful links about DeepSeek-V3.2-Exp
If you want to download the model, you have it in hugging faceThe whitepaper is on GitHub, along with implementation details and evaluations.
For comparison tests between V3.2-Exp and V3.1-Terminus, see the Official guide. And if you want to send suggestions, you have a channel of feedback public at https://feedback.deepseek.com/dsa.
With V3.2-Exp, DeepSeek prioritizes a simple idea: efficiency without sacrificing qualityDSA paves the way for models that support massive contexts at a reasonable cost, the streamlined API brings these capabilities to more teams, and the openness of the stack (weights, kernels, and documentation) makes it easier for the community to research, compare, and build real products without friction.