DeepSeek V4.1 Flash: API, Pricing, and Setup for Hermes & OpenClaw

What was released?

DeepSeek announced V4.1 Flash on September 10, 2026. Its official announcement describes a new Causal-Encoder-Decoder architecture: a 552B-parameter mixture-of-experts model, with 8B parameters active for input and 16B for output. It combines text and visual understanding. This is a new architecture, unlike the earlier V4 Flash 0731 post-training refresh.

The same announcement reports a smaller KV cache, requiring one quarter of the previous generation's HBM and one eighth of its SSD storage. These are DeepSeek's infrastructure claims, not memory requirements for your agent server. The release links to model weights and the technical report.

The API name and migration dates

Source: DeepSeek's model-name migration notes. These rules describe DeepSeek's own endpoint. An OpenRouter model, an OpenCode plan, and a direct DeepSeek key are different routes; a familiar display name alone does not identify the backend version.

DeepSeek V4.1 Flash pricing

Official direct-API USD rates per one million tokens, checked September 10. The announcement says the new prices take effect at noon Beijing time that day.

Peak hours are Monday–Friday, 01:00–04:00 and 06:00–10:00 UTC; all other hours are off-peak. The official pricing page lists a 1M context window and up to 384K output tokens, with thinking and non-thinking modes, tool calls, JSON output, Responses API, Anthropic compatibility, and vision.

For example, 1M uncached input tokens plus 100K output tokens costs $0.21 off-peak or $0.42 at peak. This is a token-budget example, not a per-task estimate: an agent may make many calls and repeatedly send context. Schedule flexible batch work off-peak, and monitor actual usage before choosing a monthly budget. Hosting and third-party provider charges are separate.

Use it on OpenClaw Launch

The dashboard uses the same model selector for Hermes Agent and OpenClaw. For a direct DeepSeek connection:

1. Create your own key in the DeepSeek API platform. See the API key guide if you are starting from scratch.
2. Open API Keys in OpenClaw Launch and save the key in the DeepSeek section.
3. Open your bot's model selector and choose the DeepSeek provider using your own key. Select deepseek-flash if it appears in the live catalog. If the catalog still lists deepseek-v4-flash, that legacy name is accepted by the direct API.
4. Start a fresh conversation so you can assess the new model without an old conversation's accumulated context. Check the selected provider as well as the model name.

Vision also depends on the client and model catalog forwarding images correctly. An old text-only Flash entry may not expose image input simply because the upstream model now supports it. For an explicit image request, use DeepSeek's vision API examples. This model understands images; it is not an image-generation endpoint.

Self-hosted Hermes Agent

Use the native DeepSeek provider and your own key. The provider and model are separate fields in Hermes; do not paste an OpenClaw-style provider prefix into model.default.

# ~/.hermes/.env
DEEPSEEK_API_KEY=YOUR_OWN_DEEPSEEK_KEY

# Merge into ~/.hermes/config.yaml
model:
  provider: deepseek
  default: deepseek-flash
  context_length: 1000000
  supports_vision: true

Use hermes model to configure the provider interactively if you prefer. Keep your existing tools, gateway settings, and other configuration. See the upstream Hermes repository and our Hermes DeepSeek guide for the surrounding setup.

Self-hosted OpenClaw

OpenClaw uses a provider/model identifier. The following explicit OpenAI-compatible provider entry avoids relying on an older built-in model catalog. Save your own DEEPSEEK_API_KEY in ~/.openclaw/.env so the gateway can read it, then merge this into your existing ~/.openclaw/openclaw.json:

{
  "models": {
    "mode": "merge",
    "providers": {
      "deepseek-direct": {
        "baseUrl": "https://api.deepseek.com/v1",
        "apiKey": "${DEEPSEEK_API_KEY}",
        "api": "openai-completions",
        "models": [{
          "id": "deepseek-flash",
          "name": "DeepSeek V4.1 Flash",
          "reasoning": false,
          "input": ["text", "image"],
          "contextWindow": 1000000,
          "maxTokens": 384000
        }]
      }
    }
  },
  "agents": {
    "defaults": {
      "model": { "primary": "deepseek-direct/deepseek-flash" },
      "models": {
        "deepseek-direct/deepseek-flash": {
          "params": {
            "extra_body": { "thinking": { "type": "disabled" } }
          }
        }
      }
    }
  }
}

This starter deliberately disables thinking: current OpenClaw DeepSeek adapters can still match only older model names when handling reasoning across tool calls. The model itself supports thinking. Verify your client version before enabling it, and preserve existing channels, model entries, and credentials. Check your configuration against the upstream provider documentation; the API's maximum context and output limits do not mean you should allocate the maximum to every agent call.

Check the direct API first

This small text request helps distinguish a provider/key issue from an agent configuration issue. Run it with your own account; it consumes API tokens. Thinking is disabled to keep this connectivity check short.

curl https://api.deepseek.com/chat/completions \
  -H "Authorization: Bearer ${DEEPSEEK_API_KEY}" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "deepseek-flash",
    "messages": [{"role": "user", "content": "Reply with OK."}],
    "thinking": {"type": "disabled"},
    "max_tokens": 32,
    "stream": false
  }'

A successful API response proves the key and endpoint work; it does not prove your bot can execute tools or process attachments. If a model is missing, check the provider and catalog. If authentication fails, check the DeepSeek key. If billing fails, check the balance on the account that owns that key.

Frequently asked questions

Choose how to run your agent

Self-hosting gives you control over the runtime, updates, and troubleshooting. Managed hosting handles the agent server while you choose the model and provider. Compare Hermes hosting and OpenClaw hosting, or read the earlier OpenClaw V4 Flash guide for historical context.

Set up a Hermes or OpenClaw agent