Yes — DeepSeek-V4-Flash-0731 is the official release of DeepSeek-V4-Flash (announced 2026-07-31, now in public beta). Here's the documentation for using it: ## 📚 Official Documentation Links **1. DeepSeek API Docs (official, primary)** - Your First API Call: https://api-docs.deepseek.com/ - Models & Pricing: https://api-docs.deepseek.com/quick_start/pricing - Using the Responses API: https://api-docs.deepseek.com/guides/responses_api - Change Log (release announcement): https://api-docs.deepseek.com/updates/ - Other guides: Thinking Mode, Tool Calls, JSON Output, Context Caching, Using the Anthropic API, Agent Integrations (all under api-docs.deepseek.com) **2. Hugging Face Model Card (official weights)** - https://huggingface.co/deepseek-ai/DeepSeek-V4-Flash-0731 - Contains: Introduction, Chat Template, How to Run with vLLM, How to Run Locally, License, Citation, Contact **3. Technical Report** - arXiv 2606.19348 — "DeepSeek-V4: Towards Highly Efficient Million-Token Context Intelligence" ## 🔑 Key Usage Facts **API calling (unchanged from preview):** - Model name: `deepseek-v4-flash` (this name now serves V4-Flash-0731 — no need to type "0731" in the API) - Base URL (OpenAI format): `https://api.deepseek.com` - Base URL (Anthropic format): `https://api.deepseek.com/anthropic` **Python example (OpenAI SDK):** ```python from openai import OpenAI client = OpenAI(api_key="<key>", base_url="https://api.deepseek.com") response = client.chat.completions.create( model="deepseek-v4-flash", messages=[{"role": "user", "content": "Hello"}], stream=False, reasoning_effort="high", extra_body={"thinking": {"type": "enabled"}} ) ``` **Specs & pricing (per 1M tokens):** - Context: 1M tokens; max output: 384K - Input (cache hit): $0.0028 | Input (cache miss): $0.14 | Output: $0.28 - Thinking mode: both non-thinking and thinking (default) supported; `reasoning_effort` supports `low` / `high` / `max` - Features: JSON output, tool calls, Responses API (native, Codex-adapted), Anthropic API, chat prefix completion, FIM completion (non-thinking only) **Responses API (for Codex):** natively supported — `client.responses.create(model="deepseek-v4-flash", instructions=..., input=...)` with base_url `https://api.deepseek.com`. Note: Responses API currently supports only `deepseek-v4-flash`, not `deepseek-v4-pro` (pro support coming early Aug 2026). **Local / self-hosted (from HF model card):** - vLLM: `vllm serve deepseek-ai/DeepSeek-V4-Flash-0731 --trust-remote-code` — enable DSpark speculative decoding with `--speculative-config '{"method":"dspark","num_speculative_tokens":7,"draft_sample_method":"greedy"}'` - SGLang: `python3 -m sglang.launch_server --model-path "deepseek-ai/DeepSeek-V4-Flash-0731"` - Transformers: `AutoModelForCausalLM.from_pretrained("deepseek-ai/DeepSeek-V4-Flash-0731", device_map="auto")` - ⚠️ No Jinja chat template — use the `encoding` folder in the HF repo (`encoding_dsv4.py`: `encode_messages`, `parse_message_from_completion_text`) to format messages - Recommended sampling: temperature 1.0, top_p 0.95 (agentic) / 1.0 (otherwise); 384K max output for high/max effort - Model: 304B params, fp8, MIT License; contact service@deepseek.com