What DeepSeek V4.1 Flash actually is
DeepSeek announced V4.1-Flash on September 9, 2026 and made it live on the API around September 10. It is not a routine refresh: DeepSeek describes it as the smallest model in a new architecture family, the first built on its Causal Encoder-Decoder (CED) design. It is a sparse mixture-of-experts model with roughly 552B total parameters, it now understands images natively (not just text), and it carries a one-million-token context window. The weights are open under the MIT License, so you can self-host or run it through third-party providers, and it is served on DeepSeek's own API under the model name deepseek-flash.
The headline specs
New architecture
First model on DeepSeek's Causal Encoder-Decoder (CED) family; a sparse MoE with ~552B total parameters.
Native multimodal
Understands image input natively, not bolted on — useful for screenshots, diagrams, and UI work.
1M-token context
A million-token window (max output up to ~393K tokens) handles very large codebases and documents in one session.
Open weights (MIT)
MIT-licensed open weights, so you can self-host or pick from many providers instead of being locked to one endpoint.
What it costs
This is the part that matters for agent coding, where token usage adds up fast. New pricing took effect September 10, 2026. On DeepSeek's own API the rates are roughly $0.15 per 1M input tokens on a cache miss, an unusually low $0.003 per 1M input tokens on a cache hit, and $0.60 per 1M output tokens. Third-party providers list similar order-of-magnitude numbers. Always check DeepSeek's current pricing page before budgeting, since rates change — but at these levels, V4.1 Flash is one of the cheapest capable coding models available.
DeepSeek V4.1 Flash pricing (per 1M tokens, DeepSeek API, Sept 2026)
| Item | Rate |
|---|---|
| Input (cache miss) | ~$0.15 |
| Input (cache hit) | ~$0.003 |
| Output | ~$0.60 |
Rates read from DeepSeek's pricing page around Sept 2026; verify the live page before you rely on them.
How to use it in Claude Code
Claude Code sends requests to whatever endpoint its config points at, and DeepSeek publishes an Anthropic-compatible endpoint. That means you can point Claude Code straight at DeepSeek — no translating proxy — and select the Flash model. You bring your own DeepSeek key and usage bills to you at the low rates above.
# Point Claude Code at DeepSeek's Anthropic-compatible endpoint
export ANTHROPIC_BASE_URL="https://api.deepseek.com/anthropic"
export ANTHROPIC_AUTH_TOKEN="your-deepseek-key"
# then run claude and select the DeepSeek Flash model (deepseek-flash)The cleaner option: pick the model, don't rewire env vars
The endpoint swap works, but you're bending Claude Code — a single-provider tool — into multi-model use. If your goal is "run DeepSeek Flash cheaply most of the time, jump to Claude for the hard stuff," it's far less friction to use an agent where DeepSeek, Claude, GLM, Kimi, Qwen and OpenAI-compatible endpoints are all just entries in a model list.
What that looks like with Wuwei
DeepSeek Flash is one entry in the list
Pick it in the UI; Claude and other providers sit right next to it.
One-click switch
Move between DeepSeek Flash and Claude mid-workflow without touching config or restarting.
Free, open source, local
The agent is free and MIT-licensed, runs on your machine, reads your real files. You only pay DeepSeek's (tiny) usage.
FAQ
When was DeepSeek V4.1 Flash released?
DeepSeek announced V4.1-Flash on September 9, 2026 and made it live on its API around September 10, 2026, with new pricing taking effect the same day.
How much does DeepSeek V4.1 Flash cost?
On DeepSeek's API the rates are roughly $0.15 per 1M input tokens (cache miss), about $0.003 per 1M cached input tokens, and $0.60 per 1M output tokens as of September 2026. That's among the lowest for a capable coding model. Check DeepSeek's live pricing page before budgeting, since rates change.
Is DeepSeek V4.1 Flash open source?
Yes. The weights are released under the MIT License, so you can self-host it or run it through third-party providers rather than being tied to a single endpoint.
Can I use DeepSeek V4.1 Flash in Claude Code?
Yes. DeepSeek exposes an Anthropic-compatible endpoint, so you set ANTHROPIC_BASE_URL to https://api.deepseek.com/anthropic and ANTHROPIC_AUTH_TOKEN to your DeepSeek key, then select the Flash model (deepseek-flash). No proxy needed. To switch between DeepSeek and Claude in a click, use a model-agnostic agent like Wuwei.