DeepSeek V4 Flash Text Generation

DeepSeek V4 Flash GA (deepseek-v4-flash-ga-260731) corresponds to DeepSeek-V4-Flash-0731, the open-source checkpoint DeepSeek promoted to general availability on July 31, 2026. The architecture matches the April preview (284B total / 13B activated MoE, 1M context) and DeepSeek states only the post-training stage was redone — yet agent benchmarks improved sharply. APIYI has completed 21 test cases plus a dedicated dual-endpoint retest; both Chat Completions and Responses are callable directly.

APIYI has launched DeepSeek V4 Flash GA: model name deepseek-v4-flash-ga-260731, available on the default / svip groups. Note that this model thinks a lot by default — pass thinking: {"type": "disabled"} explicitly for simple tasks (see “Thinking control” below).

Key Advantages

1M context that holds up

Hard input ceiling of 1,048,570 tokens. A 322K-token needle-in-a-haystack test returned in 14.77s with a correct hit; max output is 393,216 tokens.

Two caching layers

Implicit cache needs no configuration and hits 99.9% on round two; Responses adds chained explicit caching that hits the entire prior context.

Concurrency without throttling

All 20 parallel requests returned 200, with wall-clock only 1.3s above a single call — suited to high-concurrency agents and batch text jobs.

Very low unit price

$0.14 input / $0.28 output per 1M tokens, matching BytePlus list price, with cache hits as low as $0.028.

Model Information

Parameter	Value
Model name	`deepseek-v4-flash-ga-260731`
Released	July 31, 2026 (preview promoted to GA)
Architecture	284B total / 13B activated, MoE
Context window	1M (measured hard ceiling 1,048,570 tokens)
Max output	384K (measured hard ceiling 393,216 tokens)
Available groups	`default`, `svip`
Endpoints	`POST /v1/chat/completions`, `POST /v1/responses`
Deep thinking	On by default and verbose; `thinking.type` can disable it
Streaming	✅ both endpoints
Function calling / tool use	✅ both endpoints
Image input	❌ text-only model
Anthropic endpoint / Claude Code	❌ not wired — use `deepseek-v4-flash` if needed

Measured Capability Matrix

Results from APIYI testing on August 5, 2026 (official claims vs actual behavior):

Capability	Official	Chat Completions	Responses
Basic chat (non-stream / stream)	✅	✅ / ✅ (TTFB 1.43s)	✅ / ✅ (TTFB 2.31s)
Function Call (two-round loop)	✅	✅	✅
Thinking toggle `thinking.type`	✅	✅ disabled / enabled / auto all work	✅ emits reasoning items
Reasoning tiers	✅	⚠️ only `minimal` is deterministic	⚠️ same
Implicit cache	✅	✅ 99.9% hit on round two	✅ 99.9% hit
Explicit cache	✅ (Responses)	—	✅ requires `previous_response_id` chaining
Structured output	❌	❌ accepted, not enforced	❌ accepted, not enforced
Online search	✅ (Responses)	—	⚠️ wired, but backend failed 6/6
MCP	✅ (Responses)	—	❌ `AccessDenied`, account-level entitlement
Image input	—	❌	❌ explicit error

Three capabilities do not match the official sheet — know these before integrating: structured output fails silently on both endpoints (returns 200 while ignoring the schema entirely — use Function Call when you need enforcement); the online search tool is wired but its backend keeps erroring and returns no results; MCP returns AccessDenied (an account-level built-in-tool entitlement, not a model limitation).

Thinking Control

This model thinks a lot by default — a one-line question like “is 9.11 bigger than 9.9” consumed 263 reasoning tokens in our tests (the sibling deepseek-v4-flash used only 44). Disable it explicitly for simple tasks:

extra_body={"thinking": {"type": "disabled"}}   # reliably off
# or
extra_body={"reasoning_effort": "minimal"}      # 0 reasoning tokens in 10/10 runs

reasoning_effort is not a monotonic ladder. Two questions × five tiers × five samples:

Tier	river median	prob median
`minimal`	0	0
`low`	956	367
`medium`	506	193
`high`	97	153
`max`	577	173

high produced less thinking than low on both questions, and within-tier variance (low ranged 150 to 1993) far exceeds between-tier differences. Only minimal is reliable — do not treat low → max as a cost dial.

Caching

Implicit cache (automatic on both endpoints)

An identical long prefix hits on the second request: a 15,634-token prefix matched 15,616 tokens (99.9%), billed at $0.028 per million tokens.

Implicit cache requires a byte-identical prefix. Keep anything variable — timestamps, random IDs, user names — at the end of the prompt, never mixed into the prefix.

Explicit cache (Responses, requires chaining)

Common mistake: resending the same long prefix twice with caching set leaves cached_tokens at 0. The correct pattern is to write on the first call, then chain with previous_response_id:

Round	Call shape	input_tokens	cached_tokens
1 (write)	`caching: enabled`	15,629	0
2	+ `previous_response_id`	15,664	15,629
3	+ `previous_response_id`	15,701	15,664
4	+ `previous_response_id`	15,738	15,701

Quick Start

import os
from openai import OpenAI

client = OpenAI(
    api_key=os.environ["APIYI_API_KEY"],
    base_url="https://api.apiyi.com/v1",
)

resp = client.chat.completions.create(
    model="deepseek-v4-flash-ga-260731",
    messages=[{"role": "user", "content": "Explain the MoE architecture in one sentence"}],
    extra_body={"thinking": {"type": "disabled"}},
)
print(resp.choices[0].message.content)

curl https://api.apiyi.com/v1/chat/completions \
  -H "Authorization: Bearer $APIYI_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "deepseek-v4-flash-ga-260731",
    "messages": [{"role": "user", "content": "Explain the MoE architecture in one sentence"}],
    "thinking": {"type": "disabled"}
  }'

import OpenAI from "openai";

const client = new OpenAI({
  apiKey: process.env.APIYI_API_KEY,
  baseURL: "https://api.apiyi.com/v1",
});

const resp = await client.chat.completions.create({
  model: "deepseek-v4-flash-ga-260731",
  messages: [{ role: "user", content: "Explain the MoE architecture in one sentence" }],
  thinking: { type: "disabled" },
});
console.log(resp.choices[0].message.content);

Need structured output? Use Function Call

response_format has no effect on this model and raises no error — the easiest trap to fall into. Tool arguments are what actually get constrained:

tools = [{
    "type": "function",
    "function": {
        "name": "submit_result",
        "description": "Submit the extracted result",
        "parameters": {
            "type": "object",
            "properties": {
                "city": {"type": "string"},
                "temp_c": {"type": "number"},
            },
            "required": ["city", "temp_c"],
        },
    },
}]

resp = client.chat.completions.create(
    model="deepseek-v4-flash-ga-260731",
    messages=[{"role": "user", "content": "Beijing is 25 degrees today"}],
    tools=tools,
)

import json
print(json.loads(resp.choices[0].message.tool_calls[0].function.arguments))

Pricing

Item	Price
Input	$0.14 / M tokens
Output	$0.28 / M tokens
Cache hit	$0.028 / M tokens

Matches the BytePlus list price, and stacks with top-up promotions to reduce your effective cost further.

On “less than 1/10th of the flagship”: the vendor’s marketing compares against V4-Pro’s preview-era $1.74 / $3.48. Against the current V4-Pro list price ($0.435 / $0.87), this model is roughly 1/3, not 1/10.

Chat Completions

OpenAI-compatible chat endpoint with an interactive playground

Responses

Responses endpoint with chained explicit caching

Launch write-up and full test data

Benchmarks, three-way speed comparison, and the traps we hit

Model pricing table

Unit prices, endpoints and groups for every model

Basics

Basic API

Image API (Official)

Video API (Official)

Multimodal Understanding API

Text Generation

Retrieval & Safety

DeepSeek V4 Flash Text Generation

Key Advantages

1M context that holds up

Two caching layers

Concurrency without throttling

Very low unit price

Model Information

Measured Capability Matrix

Thinking Control

Caching

Implicit cache (automatic on both endpoints)

Explicit cache (Responses, requires chaining)

Quick Start

Need structured output? Use Function Call

Pricing

Chat Completions

Responses

Launch write-up and full test data

Model pricing table

​Key Advantages

1M context that holds up

Two caching layers

Concurrency without throttling

Very low unit price

​Model Information

​Measured Capability Matrix

​Thinking Control

​Caching

​Implicit cache (automatic on both endpoints)

​Explicit cache (Responses, requires chaining)

​Quick Start

​Need structured output? Use Function Call

​Pricing

​Related Pages

Chat Completions

Responses

Launch write-up and full test data

Model pricing table

Key Advantages

Model Information

Measured Capability Matrix

Thinking Control

Caching

Implicit cache (automatic on both endpoints)

Explicit cache (Responses, requires chaining)

Quick Start

Need structured output? Use Function Call

Pricing

Related Pages