DeepSeek-V4-Flash

POST /v1/chat/completions model: DeepSeek-V4-Flash

About the model

The weights served here are DeepSeek-V4-Flash-0731, DeepSeek’s official release of DeepSeek-V4-Flash, superseding the earlier preview. The vendor positions it as an agentic model: on the benchmarks published in its model card it beats DeepSeek-V4-Pro (Preview) despite a far smaller activated parameter count. The accompanying technical report is titled DeepSeek-V4: Towards Highly Efficient Million-Token Context Intelligence (arXiv:2606.19348).

Architecture

From the shipped config.json:

Architecture	DeepseekV4ForCausalLM (deepseek_v4)
Layers	43
Hidden dimension	4,096
Attention	64 query heads, 1 KV head — MLA, q_lora_rank 1,024, 64-dim RoPE split
Mixture of Experts	256 routed + 1 shared, 6 routed activated per token, expert intermediate 2,048
Vocabulary	129,280
Native context	1,048,576
Multi-token prediction	1 layer
Quantisation	FP8 e4m3, block [128, 128], dynamic activation scaling
Licence	MIT

The million-token window is native — it is max_position_embeddings in the config, not a RoPE extension applied at serve time.

What the vendor recommends

The model card suggests temperature = 1.0, with top_p = 0.95 for agentic use and top_p = 1.0 otherwise. Both parameters are accepted here, so you can follow that advice as written.

No Jinja chat template

Unlike most open-weight releases, this one ships no Jinja chat template — prompt assembly lives in a Python encoding/ folder instead. That is why this model has no opinion about where a system message sits: there is no template to raise on it. The Qwen3.8-Flash-Next page shows what happens when a model does ship one.

On this endpoint

Everything below was measured against the live endpoint. Where it disagrees with the model card, the endpoint wins — the gateway validates and rebuilds requests before they reach the backend.

At a glance

Context length	1,048,576 tokens
Input modalities	text only
Streaming	✅
Tool calling	✅ (no parallel calls)
JSON output	✅ json_object · ❌ json_schema
Thinking	✅ — off unless you ask
Stability	experimental

Thinking

Omitting reasoning_effort means no thinking. A plain request returns an empty reasoning and reasoning_tokens: 0.

The model card describes three levels — low, high, max. This endpoint accepts six values, all with 200:

minimal	low	medium	high	xhigh	max
200	200	200	200	200	200

They do not map to six distinct behaviours: minimal/low/medium produce a similar amount of thinking and high/max think noticeably longer — two effective tiers, broadly consistent with the three the vendor documents.

Where the output lands:

Thinking text	choices[0].message.reasoning
Token count	usage.completion_tokens_details.reasoning_tokens
Also available	usage.reasoning_tokens — this model emits the top-level field

The field name shifts when thinking is off

With reasoning_effort set, message contains reasoning. On a plain request with no thinking, the key present is reasoning_content instead — empty. Read reasoning and treat a missing key as “did not think”; do not branch on which key exists.

messages

Every shape we tested is accepted:

Shape	Status
system first, then user	200
system after a user turn	200
system last	200
Two system messages	200
developer instead of system	200

That is not true of Qwen3.8-Flash-Next. If one code path has to serve both, write to that model’s stricter rules.

Limits and refusals

What you send	What comes back
max_tokens beyond the window	400 Requested token count exceeds the model's maximum context length of 1048576 tokens.
response_format: json_schema	400 Model DeepSeek-V4-Flash does not support JSON schema output mode
An image_url content part	400 Model DeepSeek-V4-Flash does not support image input.
thinking: {...}	400 — use reasoning_effort

Example

curl https://developer.amd.com.cn/radeon/api/v1/chat/completions \
  -H "Authorization: Bearer $RADEON_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "DeepSeek-V4-Flash",
    "reasoning_effort": "low",
    "temperature": 1.0,
    "top_p": 0.95,
    "messages": [
      { "role": "system", "content": "Answer in one sentence." },
      { "role": "user", "content": "Why is the sky blue?" }
    ]
  }'