DSpark
DSpark is not a model. It is the speculative-decoding layer around V4: a draft path proposes tokens, the main model verifies them in parallel, and latency drops.
Speed figures come from published DeepSeek material and this site’s Features page. Self-host flags follow each runtime’s docs and can change.
What DSpark is
Speculative decoding uses a cheaper draft model to guess a stretch of tokens, then the main V4 model checks that stretch in one step. Hits keep the guessed tokens; misses roll back. DSpark tunes this loop for V4’s MoE routing — it is why hosted replies can move from “feels like 2.5 seconds” toward a few hundred milliseconds in the published illustrations (about 45 tok/s to 83 tok/s on a code-generation example, marketed as up to ~85% faster).
You never pass a dspark model ID. On https://api.deepseek.com/v1 it is on by default; public examples do not add an extra parameter. After you serve Flash or Pro from Hugging Face weights, you may need an explicit enable in that runtime (community notes: dspark enable). Treat those flags as documentation, not as an SLA.
When it matters
Watch DSpark when latency is part of the product, not a nice-to-have:
- Interactive chat and coding agents: a 2-second first token feels broken.
- High-QPS APIs: extra tok/s shortens queues and can cut GPU or token cost.
- Comparing “API default speed” with a naive local generate loop that has no draft path.
- Self-hosting Flash or Pro and wondering why the box is slower than platform.deepseek.com.
What DSpark is not
Do not put it in the same bucket as models or agent shells:
- Not a third V4 model. There is no deepseek-dspark ID.
- Not a replacement for V4 Pro or Flash. It only accelerates whichever model you already call.
- Not Harness, Deep Code, or Claude Code. Those are tools that sit on the API; DSpark sits under the API.
- Not a guarantee of 85% on your workload. Measure on your own traces.
Published-style snapshot (illustrative)
- What it is
- Speculative decoding for V4
- Official API
- On by default
- Self-host
- Enable in your runtime
- Headline boost
- Up to ~85% faster replies
- Example
- ~45 → ~83 tok/s (code gen, Features page)
- Model ID
- None — uses Pro or Flash
How to use it
On the hosted API you are probably already using it:
- Hosted: call https://api.deepseek.com/v1 with deepseek-v4-pro or deepseek-v4-flash as usual.
- Do not look for a dspark=true query param in the public OpenAI-compatible examples.
- Local: download weights from Hugging Face, serve with vLLM / SGLang / your stack, then follow that runtime’s speculative-decoding or DSpark flag.
- Benchmark first-token and tok/s on your prompts. Do not copy marketing numbers into a contract.
DSpark vs models vs tools
Accelerator, not a door. Pick the layer you actually need:
| Layer | Role | When to open this page |
|---|---|---|
| DSpark | Speed layer around V4 | You care about latency or local vs API speed |
| V4 Pro / Flash | The models being accelerated | You still need to pick a model ID |
| DeepSeek API | Hosted endpoint, DSpark on by default | You are shipping a product |
| Local deploy | Your GPU; you may have to turn DSpark on | Data cannot leave the building |
| Features page | Pro vs Flash matrix and latency graphic | You want the side-by-side numbers |
Boundaries and traps
- Self-host without the flag and you are measuring a different stack than the official API.
- Flash vs Pro still matters more than DSpark for quality. Acceleration does not add vision or 1M context.
- Agent tools (Harness, Deep Code, Claude Code) inherit whatever the endpoint does — you do not configure DSpark inside those CLIs.
- If a runtime rename the flag, the Hugging Face card and that project’s README win over this page.
Related guides
FAQ
Is DSpark a new DeepSeek model?
No. There is no dspark model ID. DSpark is speculative decoding around V4 Pro and V4 Flash. You still call deepseek-v4-pro or deepseek-v4-flash. The official API turns the accelerator on by default; self-host runtimes may need their own enable flag.
Do I pay extra for DSpark on the API?
Treat it as part of the hosted API unless the official price page lists a separate line item. You pay for Pro or Flash tokens. Self-host cost is your GPU and the draft-model overhead, not a DeepSeek invoice.
Can I turn DSpark off?
Hosted defaults are on, and public SDK examples do not show a kill switch. For local serving, use the runtime’s speculative-decoding flag. Do not assume every OpenAI-compatible proxy exposes an off switch.
Why is my local vLLM slower than the official API?
The platform stack includes DSpark. A bare vllm serve of the same weights often does not. Enable that runtime’s DSpark / speculative-decoding option, then compare first-token latency on the same prompt. Hardware and batch size still dominate.
Does DSpark replace V4 Flash when I need speed?
No. Flash is a smaller, cheaper model. DSpark makes whichever model you picked answer faster. For everyday volume pick Flash; for hard reasoning pick Pro; DSpark can sit on both.
Is the 85% figure a guarantee?
No. It is a published-style illustration (Features page: roughly 2.5s → 380ms, or 45 → 83 tok/s on a code example). Your latency depends on prompt length, batching, and GPU. Measure before you put a number in an SLA.
Where should I go if something breaks?
Hosted: official API docs and status. Self-host: the runtime README and the Hugging Face model card. This page is a community map, not the issuer of flags or keys.