DeepSeek V4.1 Flash Is Free in OpenCode: 5-Min Setup

You can put a 552-billion-parameter coding model to work this afternoon without a credit card, an API key, or a billing page. OpenCode's Zen provider currently lists DeepSeek V4.1 Flash Free at zero dollars per million input tokens, zero per million output tokens, and zero for cached reads. The listing says "for a limited time" and publishes no end date.

The catch is not hidden, but it is easy to miss. The free variant only answers inside the OpenCode app, its context window is smaller than the paid model's, and the separate $10 Go subscription meters the same model as paid capacity. Here is the setup, with the limits written down before you hand it a weekend refactor.

What the free tier actually includes

DeepSeek released V4.1 Flash on September 10, 2026. It is a mixture-of-experts model that activates roughly 8 billion parameters while reading input and 16 billion while writing, accepts text and images, and supports a 1M-token context on DeepSeek's own API. The weights sit on Hugging Face under the MIT license, though at 552B total parameters this is not a laptop model.

The free OpenCode entry is a different product surface from the API. Four limits matter:

• Price: $0 input, $0 output, $0 cached read. No account was needed in our test.
• Context: secondary write-ups put the free endpoint at around 200K tokens with 32K output, against 1M and 384K on the paid route. OpenCode's Zen docs do not state the cap, so measure it yourself before trusting a large repo to it.
• Lock-in: the model only works through the OpenCode client. Calling the Zen endpoint with curl returns a 403 FreeTierError.
• Privacy: the Zen page states the provider follows a zero-retention policy and does not train on your data. That is better than several other free models on the same list, but read the current page rather than this article before you paste in proprietary code.

Set it up in five minutes

1. Install the client

Pick one: curl -fsSL https://opencode.ai/install | bash, npm install -g opencode-ai, or brew install anomalyco/tap/opencode. On Windows the docs point you at WSL first.

2. Open the project and initialize it

Change into the repository, run opencode, then run /init once. That writes an AGENTS.md file describing your codebase so the agent stops guessing about your conventions.

3. Pick the free model

Type /models and choose DeepSeek V4.1 Flash Free. If you have not connected a provider key, the picker tends to show only the free Zen entries, which makes the right one easy to spot.

4. Make it the default, optionally

Drop this into opencode.json at the project root:

{
  "$schema": "https://opencode.ai/config.json",
  "model": "opencode/deepseek-v4.1-flash-free"
}

Note the dot in v4.1 and the -free suffix. Both are load-bearing. A hyphen or a missing suffix will throw an Unexpected server error that looks like an outage.

5. Plan before you build

Press Tab to switch into Plan mode, which proposes edits without touching files, then Tab back to Build when the diff reads sensibly. For scripted runs, skip the TUI entirely:

opencode run -m opencode/deepseek-v4.1-flash-free "fix the failing auth test" < /dev/null

The stdin redirect is not decoration. Without it, non-interactive runs can sit silent for minutes waiting for input that never arrives.

When it misbehaves

1. 403 FreeTierError: you are calling the Zen endpoint from outside OpenCode. The free model is app-only by design. Move the call into opencode run or attach a paid key.
2. "UnknownError: Unexpected server error": usually a wrong model ID, or the paid ID used without a connected Zen key. Check the dot and the -free suffix first.
3. Silent hang in scripts: add < /dev/null to the run command.
4. Context fills fast on a big monorepo: at 200K the free route runs out long before the paid one. Point the agent at specific files, restart sessions per task, or accept the paid tier.

The exit plan matters more than the setup

No end date is published, and OpenCode rotates its free list often enough that planning around the free entry is a bad habit. When the window closes, the paid routes are known. DeepSeek's own API charges deepseek-flash at $0.30 input and $1.20 output per million tokens at peak, with cached reads at $0.006; off-peak cuts every figure in half. OpenCode Zen matches those numbers from a prepaid balance. The Go plan costs $10 a month, or $40 for Go Plus, and includes $80 or $160 of V4.1 Flash usage respectively, metered across five-hour, weekly, and monthly windows.

One pricing detail is worth more than the headline numbers. DeepSeek says the KV cache for V4.1 Flash needs a quarter of the GPU memory of the previous version, and coding agents re-read enormous contexts on every tool call. Cached input is where agent bills live, which is exactly why the cached-read price is the smallest figure on the page.

Use the free tier for what it is good at: learning the agent workflow, side projects, and the moment you want to prove the thing can fix your function before you pay anyone anything. For client work, CI pipelines, or anything that needs the full million-token window, buy the paid route and schedule the heavy jobs off-peak. The free model is a trial with unusually good manners, not a budget line you can plan a quarter around.