Qwen3.8-Flash-Next

Qwen3.8-Flash-Next

Quick answer: Qwen3.8-Flash-Next is Alibaba's open-weight preview of the architecture that will underpin Qwen4 — a 125B-parameter MoE with only 6B activated per token, paired with a novel 51B-parameter n-gram embedding table (not MoE experts) for cheap parameter scaling. Released under a custom Qwen Community License (not Apache), it natively supports a 262,144-token context window (extensible to 1M via YaRN) and scores 91.7% on GPQA Diamond, 91.9% on LiveCodeBench v6, and 58.7% on DeepSWE 1.1.

At a Glance

Where Qwen3.8-Flash-Next leads

• 91.9% LiveCodeBench v6 and 91.7% GPQA Diamond — near/above-flagship reasoning and coding despite only 6B activated parameters per token
• Novel n-gram embedding design (51B params, indexed by bigrams/trigrams rather than computed) scales parameters cheaply without adding inference compute — more offload-friendly than adding MoE experts
• Qwen Sparse Attention (QSA) operates at the micro-block level rather than per-token, cutting long-context latency for agentic workloads
• Outperforms the much larger dense Qwen3.8-27B across nearly every benchmark reported, despite only 6B active parameters vs. 27B dense

Where it lags

• Custom Qwen Community License 1.0 — not a plain permissive license like Apache 2.0; check the exact terms before commercial deployment
• Native context (262,144 tokens) is smaller than GLM-5.3-Flash's native 1M — Qwen3.8-Flash-Next needs YaRN RoPE scaling to reach 1M
• Explicitly labeled an experimental architecture preview, not a production flagship — the hosted "Qwen3.8-Flash" API sibling is the production-hardened version

Best for: Teams evaluating next-generation efficient-MoE architectures (n-gram embeddings + hybrid linear/sparse attention) for coding and agentic workloads, or wanting an early, self-hostable look at Qwen4-generation techniques.

What Qwen3.8-Flash-Next Is

Qwen3.8-Flash-Next is Alibaba's Qwen Team's first open-weight release built on an experimental architecture positioned as a preview of what will underpin Qwen4. Rather than iterating on parameter count, the release is explicitly framed around architectural efficiency: "the question is no longer just how much we can scale, but how efficiently we can do so."

The model has 125B total parameters with only 6B activated per token via MoE routing (512 experts, 10 routed + 1 shared per token), plus a separate 51B-parameter n-gram embedding table — a novel scaling axis that indexes 20 million bigrams/trigrams at layer 2 rather than adding compute-heavy MoE experts, making it more amenable to memory-constrained/offloaded serving. Attention is a hybrid design across 48 layers: most blocks use Gated DeltaNet (linear attention), with a smaller share using Qwen Sparse Attention (QSA) — a rework of the prior Gated Attention that reasons over micro-blocks rather than individual tokens, cutting long-context latency. A Gated Residual mechanism (4 branches, element-wise read/write gates) adds finer per-layer expressiveness without much inference overhead. Training used a tailored Muon+AdamW recipe that skips traditional batch-size warmups.

Qwen3.8-Flash-Next natively supports a 262,144-token context window, extensible to 1,000,000 tokens via YaRN RoPE scaling (recommended only when genuinely needed, since static YaRN can hurt short-context performance). It thinks by default (enable_thinking=True), supports a reasoning_effort parameter (xhigh/medium/low), and retains thinking blocks across turns via preserve_thinking for agentic consistency. Alibaba also offers Qwen3.8-Flash, a production-hardened hosted version via Qwen Cloud with a 1M-token context by default and built-in tools — no independently published benchmark table exists for that hosted variant separate from Flash-Next, so it isn't tracked as a separate Benchgen page.

Specifications

Pricing

Open weights available on Hugging Face (Qwen/Qwen3.8-Flash-Next) — free to self-host under the Qwen Community License 1.0 (verify commercial-use terms). A production-hardened hosted version, Qwen3.8-Flash, is available via Qwen Cloud — check Qwen Cloud's pricing page for current rates (not independently verified here).

Context Window

Qwen3.8-Flash-Next natively supports 262,144 tokens (256K), extensible to 1,000,000 tokens (1M) via YaRN RoPE scaling on supported serving frameworks (vLLM, SGLang, TokenSpeed). Alibaba recommends only enabling YaRN when long-context workloads genuinely require it, since static YaRN scaling can slightly reduce short-context performance.

Public Benchmark Scores

Scores are self-reported by Alibaba's Qwen Team on the official Hugging Face model card, cross-checked against the raw HTML benchmark tables (which include explicit column headers identifying comparison models) rather than the stripped structured Eval Results metadata alone. Agents' Last Exam is reported by Qwen as two metrics ("Pass@1" 24.3 and "Score" 51.2); Benchgen's leaderboard tracks the "Score" metric (converted to the platform's 0–1 fractional scale as 0.512), consistent with the existing Qwen3.8-27B/Qwen3.8-Max entries on this benchmark.

Additional benchmarks reported by Qwen, not yet on Benchgen's leaderboard:

• CoWorkBench (73.9%) and RecreationBench (49.9%) are explicitly labeled in-house/proprietary Qwen evaluations — not independently reproducible.
• AndroidWorld (84.5%), ClawEval-MM (Pass@3 64.4% / Average 60.4%), and ERQA (72.3%) and LVBench (76.6%) appear to be real, independently-documented benchmarks but don't have Benchgen pages yet (same gap already flagged on the Qwen3.8-27B page) — follow-up opportunity.
• Vision2Web (64.0%) is a modified/non-standard variant of existing web-agent benchmark families without clear independent documentation of the exact modifications — not added pending further verification (same skip reason as Qwen3.8-27B).
• OSWorld 2.0 (reported as Binary 19.4% / Partial 52.3%) — Benchgen's existing osworld-2 leaderboard entries (e.g. Kimi K3 at 58.3%) don't document which of these two sub-metrics they represent, so this score is skipped pending clarification rather than risking a scale mismatch.

Qwen3.8-Flash-Next vs Alternatives

Despite activating only 6B parameters per token — far fewer than the 27B-dense Qwen3.8-27B — Qwen3.8-Flash-Next beats it on every directly-compared metric, underscoring the efficiency case for the hybrid linear/sparse-attention + n-gram-embedding architecture over simply scaling a dense model.

Related Models & Benchmarks

• Compare: Qwen3.8-27B · Qwen3.8 Max · GLM-5.3-Flash
• Benchmarks: DeepSWE · GPQA Diamond · LiveCodeBench v6 · Agents' Last Exam · Humanity's Last Exam

Use via API

from openai import OpenAI

client = OpenAI(base_url="http://localhost:8000/v1", api_key="EMPTY")

completion = client.chat.completions.create(
    model="Qwen/Qwen3.8-Flash-Next",
    messages=[{"role": "user", "content": "Write a Python function to merge two sorted linked lists."}],
    extra_body={"chat_template_kwargs": {"enable_thinking": True}},
    reasoning_effort="xhigh",
)

Frequently Asked Questions

Specs from Alibaba's official Hugging Face model card (huggingface.co/Qwen/Qwen3.8-Flash-Next). Last updated 2026-08-30.