Short answer: DeepSeek-V3.1 is a hybrid model with thinking and non-thinking modes, documented tool-calling support, and a model-card context length of 128K. The model card lists 671B total parameters and 37B activated parameters. Those are specifications for the V3.1 model release—not a promise that the same model or limits are available through today’s hosted API. DeepSeek announced that the former deepseek-chat and deepseek-reasoner names would be retired after July 24, 2026. As of November 9, 2026, DeepSeek’s API Quick Start uses deepseek-flash; its pricing page maps that API name to DeepSeek-V4.1-Flash-1017. That is a current API model, not a V3.1 identifier. (retirement announcement; API pricing and model version)
DeepSeek-V3.1 at a glance
DeepSeek’s V3.1 release announcement describes one model with two inference modes and says post-training improved tool use and multi-step agent tasks. At launch, the announcement mapped deepseek-chat to non-thinking mode and deepseek-reasoner to thinking mode, and stated 128K context for both. Those are historical launch details, not the current API contract.
The V3.1 model card lists 671B total parameters, 37B activated parameters, and 128K context. Some third-party pages claim 685B total parameters; that conflicts with the figure in the model card, so this guide uses the model-card figure rather than combining the claims. The listed context length describes the model-card release and does not establish limits for every serving route.
Thinking mode, non-thinking mode, and tools
The V3.1 model card describes mode selection through the chat template. In its first-turn local template, non-thinking mode uses an assistant prefix ending with </think>; thinking mode uses one ending with <think>. This is a tokenizer/template distinction, not a universal instruction for every API provider.
The model card says tool calling is supported in non-thinking mode. It also describes using a user-provided search tool in a multi-turn process for questions needing external or current information. That means V3.1 can be connected to a search tool; the model card does not establish that V3.1 itself includes built-in live web search. Its tool-call format requires a tool name matching an available tool and JSON arguments matching that tool’s schema. (V3.1 model card)
Practical starting point: evaluate non-thinking mode for ordinary chat and structured tool use; test thinking mode for tasks where additional reasoning may help. Compare both using your own prompts. The release announcement’s faster-thinking comparison with DeepSeek-R1-0528 is a company-reported claim, not a latency guarantee for your application. (V3.1 release announcement)
What do the benchmarks show?
The V3.1 model card reports different scores by task and mode. For example, it reports LiveCodeBench (2408–2505), Pass@1 at 56.4 for V3.1-NonThinking and 74.8 for V3.1-Thinking. For AIME 2025, Pass@1, it reports 49.8 and 88.4, respectively. These are model-card results, not independent tests or predictions of production performance.
Evaluation conditions matter: the model card says search-agent tests used an internal framework with a commercial search API, a webpage filter, and a 128K context window; SWE-bench used an internal code-agent framework. Agent benchmark outcomes therefore reflect more than the base model alone. (V3.1 model card)
Use benchmark results to choose what to test, not as a universal ranking. Build a small set of representative requests and score correctness, completion rate, latency, tool-call accuracy, and failure behavior. Test the exact mode and serving route you intend to deploy.
Production API: what changed
The V3.1 model card shows local Transformers tokenizer and chat-template use; that is not the same as a hosted HTTP API integration. The V3.1-era API names deepseek-chat and deepseek-reasoner had an announced retirement deadline of July 24, 2026. That deadline has passed. DeepSeek’s current Quick Start uses deepseek-flash; the pricing page maps it to DeepSeek-V4.1-Flash-1017. This is a current API offering, not a V3.1 model identifier.
For a new integration with DeepSeek’s hosted API, its current Quick Start documents an OpenAI-compatible base URL, Bearer-token API-key authentication, the POST /chat/completions endpoint, and deepseek-flash as a model name. The example below is for that current API model—not V3.1.
export DEEPSEEK_API_KEY="your_api_key"
curl https://api.deepseek.com/chat/completions \
-H "Content-Type: application/json" \
-H "Authorization: Bearer ${DEEPSEEK_API_KEY}" \
-d '{
"model": "deepseek-flash",
"messages": [
{"role": "user", "content": "Hello!"}
],
"stream": false
}'
DeepSeek’s pricing page lists model prices per 1 million tokens and bills input and output tokens. Its current deepseek-flash listing gives these rates:
| Token type | Peak | Off-peak |
|---|---|---|
| Input, cache miss | $0.30 / 1M tokens | $0.15 / 1M tokens |
| Input, cache hit | $0.006 / 1M tokens | $0.003 / 1M tokens |
| Output | $1.20 / 1M tokens | $0.60 / 1M tokens |
The pricing page defines peak hours as 01:00–04:00 and 06:00–10:00 UTC, Monday through Friday, excluding Chinese public holidays. All other times—including weekends and Chinese public holidays in full—are off-peak. These are listed prices for the current deepseek-flash API model, not V3.1 prices; DeepSeek says prices may change, so check the live pricing page when budgeting. (DeepSeek API pricing)
Should you choose DeepSeek-V3.1?
V3.1 is a reasonable evaluation candidate if you need to assess its released weights, hybrid prompting behavior, or published task-specific results. Its model card states that the repository and weights are licensed under MIT; check the license attached to the exact artifact you plan to use. For a new production integration with DeepSeek’s hosted API, choose from the currently documented API model names rather than planning around V3.1’s retired API names. (retirement announcement)
Before committing, confirm:
- The exact model identifier and its availability on your chosen route.
- Context and output limits for that model and interface.
- Whether the intended mode and tool-calling behavior are supported.
- Quality, latency, and tool reliability on representative tasks.
- Current cost using expected input/output token volumes and cache-hit rates.
- For self-hosting, the hardware, serving stack, license, and operating cost for the exact weights.
In short: DeepSeek-V3.1 is a model release with published weights and evaluation details; DeepSeek’s current hosted API uses a different model name and serves a newer model. Keep those choices separate when selecting a model or planning an integration.
Sources
DeepSeek API Legacy-Model Retirement Announcement