Canonical names (recommended)
Write your code against these:| Canonical name | What you get |
|---|---|
deepseek-v4-flash | DeepSeek V4 Flash — fast non-thinking mode, default for general chat / completion |
deepseek-v4-pro | DeepSeek V4 Pro — adds the thinking / reasoning mode (set reasoning_effort to control depth) |
Legacy compat (until 2026-07-24)
DeepSeek announced these names will be removed on 2026-07-24. ByteSpike’s gateway will continue to accept them until then, resolving each request to the canonical name below.| Legacy name | Resolves to | Note |
|---|---|---|
deepseek-chat | deepseek-v4-flash | The most common pre-V4 DeepSeek string |
deepseek-reasoner | deepseek-v4-pro (thinking mode) | The pre-V4 “reasoner” handle |
Names not to use
Earlier DeepSeek lineages —deepseek-coder, deepseek-v2 family handles — have been retired and are no longer accepted by ByteSpike. Requests using them return a 404 with a hint pointing at the V4 line. (Note: deepseek-r1 and deepseek-v3-2 are still served as canonical models — they are not legacy; see the models index.)
For other model families, ByteSpike accepts the canonical model ID directly. No alias layer applies; there are no synonyms to memorize. Look up the exact string for each model on its models page.
How to migrate
A one-line code change in most clients:deepseek-reasoner and depend on the thinking output specifically, also confirm your client sets reasoning_effort (default is medium on V4 Pro).
Next
- Models index — every canonical model ByteSpike serves, with pricing and protocol coverage
- Endpoint types — the protocols each model can be called through