DeepSeek: DeepSeek V4.1 Flash
3 providers live- 3 live providers
- 100% uptime
- 0.40 s first token
- 193 tok/s output
DeepSeek V4.1 Flash is a sparse mixture-of-experts model from DeepSeek, and the first built on the company's Causal Encoder-Decoder (CED) architecture. It activates 8B parameters on input and 16B on output from a 552B-parameter backbone, an asymmetric split that keeps per-token compute low relative to the model's total size. Image understanding is native to the architecture, with visual and text embeddings trained jointly from the start of pre-training rather than added afterward as in the earlier experimental V4 Flash Vision Exp. It is suited for coding, terminal, and computer-use agents, along with long-horizon tasks that must run to completion across many steps and long-context analysis. Compressed KV caching cuts cache memory to roughly a quarter of the previous Flash generation, significantly reducing costs on agentic workloads. DeepSeek positions it as the cost-efficient tier of the V4.1 family and reports that it exceeds V4 Pro on performance, speed, and task completion time.
- Context
- 1M
- tokens
- Max output
- 128K
- tokens
- Input
- $0.30
- per million tokens
- Output
- $1.20
- per million tokens
Providers
Ranked by price and speed. Requests go to the best healthy provider, then fail over down the list.Benchmarks| Who answers now | Input per 1M | First token |
|---|---|---|
 Prime Intellectpicked | $0.30 | 0.48 s |
 DeepInfra | $0.20 | 0.58 s |
 ScalewayEU | $0.6915 | 0.40 s |
Prime Intellect answers now: first in the router's order.
First token is the 24-hour median. If the pick fails, the next one answers at its own price.
| # | Provider | Input | Output | Cache read | Latency | Throughput | Uptime (3d) |
|---|---|---|---|---|---|---|---|
| 1 | Data policy not declared | $0.30 | $1.20 | $0.006 | 0.48 s | 193 tok/s | 100% |
| 2 | Data policy not declared | $0.20 | $0.60 | $0.006 | 0.58 s | 78 tok/s | 100% |
| 3 | Data policy not declared | $0.6915 | $2.0743 | $0.1383 | 0.40 s | 161 tok/s | 95.8% |
Price history
Input price since Oct 20, 2026, in USD per million tokens.
- DeepInfra
- Prime Intellect
- Scaleway
Data table
| Provider | From | Input | Output |
|---|---|---|---|
| DeepInfra | Oct 20, 2026 | $0.20 | $0.60 |
| Prime Intellect | Oct 21, 2026 | $0.30 | $1.20 |
| Scaleway | Oct 21, 2026 | $0.6931 | $2.0793 |
| Scaleway | Oct 22, 2026 | $0.6943 | $2.0828 |
| Scaleway | Oct 22, 2026 | $0.6925 | $2.0775 |
| Scaleway | Oct 23, 2026 | $0.6918 | $2.0752 |
| Scaleway | Oct 26, 2026 | $0.6915 | $2.0743 |
Performance
Last 24 hours, from llm.buzz's test requestsThroughput (tok/s)
- Prime Intellect193 tok/s
- DeepInfra78 tok/s
- Scaleway161 tok/s
Data table
| Provider | From (UTC) | p50 | p90 | p99 |
|---|---|---|---|---|
| Prime Intellect | Oct 26, 05:00 | 193 tok/s | 193 tok/s | 193 tok/s |
| Prime Intellect | Oct 26, 10:00 | 134 tok/s | 134 tok/s | 134 tok/s |
| Prime Intellect | Oct 26, 15:00 | 193 tok/s | 193 tok/s | 193 tok/s |
| Prime Intellect | Oct 26, 19:00 | 161 tok/s | 161 tok/s | 161 tok/s |
| Prime Intellect | Oct 26, 23:00 | 232 tok/s | 232 tok/s | 232 tok/s |
| Prime Intellect | Last 24 hours | 193 tok/s | 134 tok/s | 134 tok/s |
| DeepInfra | Oct 26, 06:00 | 78 tok/s | 78 tok/s | 78 tok/s |
| DeepInfra | Oct 26, 11:00 | 112 tok/s | 112 tok/s | 112 tok/s |
| DeepInfra | Oct 26, 15:00 | 45 tok/s | 45 tok/s | 45 tok/s |
| DeepInfra | Oct 26, 19:00 | 65 tok/s | 65 tok/s | 65 tok/s |
| DeepInfra | Oct 26, 23:00 | 93 tok/s | 93 tok/s | 93 tok/s |
| DeepInfra | Oct 27, 03:00 | 54 tok/s | 54 tok/s | 54 tok/s |
| DeepInfra | Last 24 hours | 78 tok/s | 45 tok/s | 45 tok/s |
| Scaleway | Oct 26, 05:00 | 161 tok/s | 161 tok/s | 161 tok/s |
| Scaleway | Oct 26, 10:00 | 112 tok/s | 112 tok/s | 112 tok/s |
| Scaleway | Oct 26, 14:00 | 112 tok/s | 112 tok/s | 112 tok/s |
| Scaleway | Oct 26, 18:00 | 193 tok/s | 193 tok/s | 193 tok/s |
| Scaleway | Oct 26, 22:00 | 161 tok/s | 161 tok/s | 161 tok/s |
| Scaleway | Oct 27, 02:00 | 232 tok/s | 232 tok/s | 232 tok/s |
| Scaleway | Last 24 hours | 161 tok/s | 112 tok/s | 112 tok/s |
Latency (time to first token)
- Prime Intellect0.48 s
- DeepInfra0.58 s
- Scaleway0.40 s
Data table
| Provider | From (UTC) | p50 | p90 | p99 |
|---|---|---|---|---|
| Prime Intellect | Oct 26, 05:00 | 0.48 s | 0.48 s | 0.48 s |
| Prime Intellect | Oct 26, 10:00 | 0.83 s | 0.83 s | 0.83 s |
| Prime Intellect | Oct 26, 15:00 | 0.58 s | 0.58 s | 0.58 s |
| Prime Intellect | Oct 26, 19:00 | 0.48 s | 0.48 s | 0.48 s |
| Prime Intellect | Oct 26, 23:00 | 0.48 s | 0.48 s | 0.48 s |
| Prime Intellect | Last 24 hours | 0.48 s | 0.83 s | 0.83 s |
| DeepInfra | Oct 26, 06:00 | 0.58 s | 0.58 s | 0.58 s |
| DeepInfra | Oct 26, 11:00 | 0.40 s | 0.40 s | 0.40 s |
| DeepInfra | Oct 26, 15:00 | 0.69 s | 0.69 s | 0.69 s |
| DeepInfra | Oct 26, 19:00 | 1.20 s | 1.20 s | 1.20 s |
| DeepInfra | Oct 26, 23:00 | 0.58 s | 0.58 s | 0.58 s |
| DeepInfra | Oct 27, 03:00 | 1.20 s | 1.20 s | 1.20 s |
| DeepInfra | Last 24 hours | 0.58 s | 1.20 s | 1.20 s |
| Scaleway | Oct 26, 05:00 | 0.40 s | 0.40 s | 0.40 s |
| Scaleway | Oct 26, 10:00 | 0.83 s | 0.83 s | 0.83 s |
| Scaleway | Oct 26, 14:00 | 0.58 s | 0.58 s | 0.58 s |
| Scaleway | Oct 26, 18:00 | 0.40 s | 0.40 s | 0.40 s |
| Scaleway | Oct 26, 22:00 | 0.33 s | 0.33 s | 0.33 s |
| Scaleway | Oct 27, 02:00 | 0.33 s | 0.33 s | 0.33 s |
| Scaleway | Last 24 hours | 0.40 s | 0.83 s | 0.83 s |
Uptime
Last 72 hours, by hour- Up: 95% or more served
- Degraded: 85% to 95%
- Down: below 85%
- No data
- From llm.buzz's test requests
Questions
How much does DeepSeek: DeepSeek V4.1 Flash cost?
DeepSeek: DeepSeek V4.1 Flash costs $0.30 per million input tokens and $1.20 per million output tokens at Prime Intellect, the provider requests go to first, with cache reads at $0.006 per million tokens. Each provider sets its own price, and you pay the price of the provider that answered.
Do providers train on my prompts?
Each provider declares it, and the table above shows every declaration. You can limit a request to providers that keep nothing, or to providers in the EU.
Can I serve DeepSeek: DeepSeek V4.1 Flash myself?
Yes. Apply as a provider, add your endpoint and list this model at your price. It goes live once the automated check passes.
Similar models
Call it
import os
from openai import OpenAI
client = OpenAI(
base_url="https://api.llm.buzz/api/v1",
api_key=os.environ["LLM_BUZZ_KEY"],
)
reply = client.chat.completions.create(
model="deepseek/deepseek-v4.1-flash",
messages=[{"role": "user", "content": "Hello"}],
)import OpenAI from "openai";
const client = new OpenAI({
baseURL: "https://api.llm.buzz/api/v1",
apiKey: process.env.LLM_BUZZ_KEY,
});
const reply = await client.chat.completions.create({
model: "deepseek/deepseek-v4.1-flash",
messages: [{ role: "user", content: "Hello" }],
});curl https://api.llm.buzz/api/v1/chat/completions \
-H "Authorization: Bearer $LLM_BUZZ_KEY" \
-H "Content-Type: application/json" \
-d '{"model": "deepseek/deepseek-v4.1-flash",
"messages": [{"role": "user", "content": "Hello"}]}'Endpoints
POST /api/v1/chat/completionsPOST /api/v1/responsesPOST /api/v1/messagesAuthentication
Send your llm.buzz key as a Bearer token: Authorization: Bearer $LLM_BUZZ_KEY. Create keys in the console. The same key calls every model; each request pays the price of the provider that answered it.
More in Authentication and the API reference.
README
DeepSeek V4.1 Flash is a sparse mixture-of-experts model from DeepSeek, and the first built on the company's Causal Encoder-Decoder (CED) architecture. It activates 8B parameters on input and 16B on output from a 552B-parameter backbone, an asymmetric split that keeps per-token compute low relative to the model's total size. Image understanding is native to the architecture, with visual and text embeddings trained jointly from the start of pre-training rather than added afterward as in the earlier experimental V4 Flash Vision Exp. It is suited for coding, terminal, and computer-use agents, along with long-horizon tasks that must run to completion across many steps and long-context analysis. Compressed KV caching cuts cache memory to roughly a quarter of the previous Flash generation, significantly reducing costs on agentic workloads. DeepSeek positions it as the cost-efficient tier of the V4.1 family and reports that it exceeds V4 Pro on performance, speed, and task completion time.
- Model ID
- deepseek/deepseek-v4.1-flash
- Takes
- Text, image
- Answers
- Text
- Context
- 1M tokens
- Max output
- 128K tokens
- Capabilities
- Tools, Caching, Structured outputs
- Parameters
- max_tokens, response_format, structured_outputs, temperature, tool_choice, tools
- Weights
- Open weights
- On llm.buzz since
- Oct 20, 2026