Qwen3.8-Flash-Next

POST /v1/chat/completions model: Qwen3.8-Flash-Next

About the model

The weights served here are Qwen3.8-Flash-Next-FP8 — the FP8-quantised build of Qwen3.8-Flash-Next, using fine-grained FP8 with block size 128. Qwen states its metrics are nearly identical to the unquantised original.

Qwen describes this release as an experimental preview of the architecture that will underpin Qwen4, and the first open-weight model built on it. Three things are new relative to the Qwen3 line, per the model card:

• Hybrid attention with QSA — the Gated DeltaNet plus Gated Attention pairing is reworked into Gated DeltaNet plus Qwen Sparse Attention.
• Gated Residual — normalised residual streams get an extra gate, aimed at keeping deep, wide models trainable.
• N-gram Embedding — a parameter-scaling axis that costs less compute than MoE and is easier to offload.

Details are in Qwen’s blog post and technical report.

This is not the same as “Qwen3.8-Flash”

Qwen also operates a managed service model called Qwen3.8-Flash on Qwen Cloud, built on these weights with additional production features — a 1M context window by default among them. What this endpoint serves is the open-weight Qwen3.8-Flash-Next, at its native 262,144-token window.

Architecture

From the model card and the shipped config.json:

Parameters	125B total, 6B activated, plus 51B n-gram embedding and 4B MTP
Layers	48, laid out as 12 × (3 × Gated DeltaNet→MoE, then 1 × QSA→MoE)
Hidden dimension	2,560
Gated DeltaNet	48 V heads, 16 QK heads, head dim 128
Qwen Sparse Attention	24 query heads, 2 KV heads, head dim 256, 64-dim RoPE
QSA indexer	MQA with 4 query heads and 1 shared key head, head dim 128
QSA budget	512 blocks / 2,048 tokens
Mixture of Experts	512 experts, 10 routed + 1 shared activated, expert intermediate 640
N-gram embedding	20,000,000 bigrams/trigrams, applied at layer 2
Gated Residual	4 branches, bottleneck rank 320
Vocabulary	248,320 (padded)
Native context	262,144, extensible to 1,000,000
Multi-token prediction	1 layer
Quantisation	FP8, block [128, 128], dynamic activation scaling
Licence	Qwen Community License 1.0

The QSA budget is the number worth remembering: attention over the KV cache is capped at 2,048 selected tokens regardless of how long the conversation is, and 36 of the 48 layers are linear attention whose state does not grow with context at all.

The model card says “with Vision Encoder” — image input is still refused here

Qwen lists the type as Causal Language Model with Vision Encoder, and the repository is tagged image-text-to-text. This endpoint rejects image content anyway — see Limits and refusals. Treat the model as text-only when calling it here.

On this endpoint

Everything below was measured against the live endpoint. Where it disagrees with the model card, the endpoint wins.

At a glance

Context length	262,144 tokens
Input modalities	text only
Streaming	✅
Tool calling	✅ (no parallel calls)
JSON output	✅ json_object · ❌ json_schema
Thinking	✅ — on unless you turn it down
Stability	experimental

messages — the two rules that bite

Exactly one system message, and it must be first

Anything else is refused:

Shape	Status
system first, then user	200
system after a user turn	400
system last	400
Two system messages	400

{
  "error": {
    "message": "System message must be at the beginning.",
    "type": "BadRequestError",
    "code": 400
  }
}

This is the model’s own Jinja chat template raising, not a gateway rule — Qwen ships the check in tokenizer_config.json. A dedicated endpoint running the same weights refuses the same requests.

The developer role is not accepted

Newer OpenAI SDKs emit developer where older ones emit system. The accepted role set is system, user, assistant, tool — developer is not in it, and it fails during request-body deserialisation, so the status is 422, not 400:

Failed to deserialize the JSON body into the target type: messages[0]: unknown role: developer

If your client library defaults to developer, override it to system.

Thinking

Omitting reasoning_effort does not disable thinking. A plain request already returns a populated reasoning and a non-zero reasoning_tokens. The model’s internal default is xhigh, its longest tier.

Only two tiers can actually be requested:

Tier	Status	Why
low	200
medium	200
high	400	Passes request validation, then the model refuses: Unexpected reasoning effort high. Supported types are xhigh (default), medium, and low.
xhigh	422	Rejected during deserialisation — not in the endpoint’s enum
max	422	Same
minimal	422	Same

xhigh is the default but cannot be requested

The model names xhigh as its default, yet the endpoint’s request schema only accepts low, medium, high — so xhigh is reachable only by omitting the parameter, and high passes the schema then gets refused by the model. Send low or medium explicitly. Those are the only two values that work end to end.

Where the output lands:

Thinking text	choices[0].message.reasoning
Token count	usage.completion_tokens_details.reasoning_tokens
Not emitted	usage.reasoning_tokens — no top-level field, unlike DeepSeek-V4-Flash

Limits and refusals

What you send	What comes back
max_tokens beyond the window	400 max_tokens=… cannot be greater than max_model_len=max_total_tokens=262144.
response_format: json_schema	400 Model Qwen3.8-Flash-Next does not support JSON schema output mode
An image_url content part	400 Model Qwen3.8-Flash-Next does not support image input.
thinking: {...}	400 — use reasoning_effort

max_tokens is capped against the total budget, prompt included — 262,144 covers input plus output, not output alone.

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": "Qwen3.8-Flash-Next",
    "reasoning_effort": "low",
    "messages": [
      { "role": "system", "content": "Answer in one sentence." },
      { "role": "user", "content": "Why is the sky blue?" }
    ]
  }'

One system message, first in the array, and an explicit low — that request shape also works unchanged on DeepSeek-V4-Flash.