On our own GPUs
DeepSeek V4 Flash API
284B total, 13B active per token — the lowest unit cost in the V4 generation.
Context
1M
Max output
384K
Serving precision
BF16
Licence
MIT
DeepSeek V4 Flash shipped alongside V4 Pro: 284B total parameters with 13B active per token, the same 1-million-token context window and the same thinking / non-thinking modes, under the MIT licence. Its positioning is unambiguous — keep the V4 generation’s capability while pushing cost per million tokens as low as it goes, so it can absorb bulk production traffic that is sensitive to both latency and price.
Open weights, deployed in facilities we build out and lease directly, operated by us. We resell nobody else’s account quota.
Specifications
- Model ID
- deepseek-v4-flash
- Published by
- DeepSeek
- How it is served
- On our own GPUs
- Licence
- MIT
- Total parameters
- 284B MoE
- Active parameters
- 13B
- Context
- 1M
- Max output
- 384K
- Serving precision
- BF16
- Input modalities
- Text
- Released
- 2026-04-24
- Price
- See pricing
What it is good at
- Only 13B active
- Sparse activation keeps the knowledge of a large model while running a small fraction of the network per token. The same hardware absorbs far more concurrency, which shows up directly in your unit price and your queue time.
- Long context was not the thing they cut
- Flash is not cheap because the window shrank — it still carries the full 1M tokens. Batch-processing long documents does not force you down to a smaller-window model.
- A good routing floor
- Most requests do not need the top tier. Flash on the main path with escalation to V4 Pro or Kimi K3 is one of the best cost/quality combinations we see in production.
Where it fits
- High-volume classification, extraction, cleaning and labelling
- High-concurrency chat and first-line support
- Summarisation of logs, reviews and tickets at scale
- The default tier in a multi-model routing setup
Three steps
Change one line of base_url, put the model ID below in the `model` field, and leave the rest of your code alone.
model: "deepseek-v4-flash"
curl https://api.nexaix.net/v1/chat/completions \
-H "Authorization: Bearer $NEXAIX_API_KEY" \
-H "Content-Type: application/json" \
-d '{
"model": "deepseek-v4-flash",
"stream": true,
"messages": [
{ "role": "user", "content": "Summarise this contract." }
]
}'import os
from openai import OpenAI
client = OpenAI(
# the only new line
base_url="https://api.nexaix.net/v1",
api_key=os.environ["NEXAIX_API_KEY"],
)
# everything below is your existing code
stream = client.chat.completions.create(
model="deepseek-v4-flash",
stream=True,
messages=[
{"role": "user", "content": "Summarise this contract."}
],
)
for chunk in stream:
print(chunk.choices[0].delta.content or "", end="")import OpenAI from "openai";
const client = new OpenAI({
baseURL: "https://api.nexaix.net/v1", // the only new line
apiKey: process.env.NEXAIX_API_KEY,
});
// everything below is your existing code
const stream = await client.chat.completions.create({
model: "deepseek-v4-flash",
stream: true,
messages: [
{ role: "user", content: "Summarise this contract." }
],
});
for await (const chunk of stream) {
process.stdout.write(chunk.choices[0]?.delta?.content ?? "");
}About this model
How much worse is Flash than Pro?
It depends entirely on the task. On extraction, classification and ordinary conversation the gap is small; on mathematics, hard algorithms and cross-file reasoning it widens sharply. Don’t take anyone’s word for it — run both against your own eval and set the routing threshold from the result.
Are there rate limits?
Every key has its own rate limit and quota, documented in the docs. At capacity you get a standard 429 with retry guidance — we never quietly hand the request to a different model.
Can one key reach both Flash and Pro?
Yes. A single key reaches every model you have enabled; change the `model` field and nothing else. Usage is metered and itemised per model in the console.
Also worth a look
- On our own GPUsKimi K32.8T mixture-of-experts, a 1M-token window, and vision built into the base.kimi-k3
- On our own GPUsGLM-5.2750B sparse MoE, a 1M window, MIT licensed — tuned hard for software engineering.glm-5.2
- On our own GPUsDeepSeek V4 Pro1.6 trillion parameters, thinking and non-thinking modes, MIT licensed.deepseek-v4-pro
Run your own eval on it
Take the trial credit, run the evals you already have against this model, and then decide whether it gets your production traffic.
