Wait Which Model?
← Back to directory
Alibaba (Qwen)

Qwen3.8-Flash-Next

Released Aug 26, 2026 · knowledge cutoff unpublished

Status
Superseded
Location
China
Modality
Multimodal
Context window
262K
Max output
131K
Speed
88 tok/s · 2.57s to first answer token
Price ($/MTok in / out)
$0.16 / $0.47
Cost per task
$0.090
Open weights
Yes
Benchmarks3 of 10 reported

GPQA Diamond

91.7%

SWE-bench Pro

62.5%

Humanity's Last Exam

35.9%

MMLU-Proretired

—

SWE-bench Verified

—

Terminal-Bench 2.1

—

AIME (latest)retired

—

LMArena Elo

—

GDPval-AA v2

—

ARC-AGI-2

—
Strengths & weaknesses

Strengths

  • Only a sliver of the weights fire per token and the n-gram table can sit in ordinary system RAM — a checkpoint this large runs from CPU memory at close to VRAM speeds on a single workstation
  • Thinking is on by default but switchable per request, with a low/medium/xhigh reasoning_effort dial and preserve_thinking to carry reasoning across agent turns instead of re-deriving it
  • Image and video understanding are native to the checkpoint — charts, documents, hour-scale video, computer-use and mobile-use tasks — not a vision adapter bolted onto a text base
  • Does its best work on open-ended agentic coding — reviewers found it stronger when given room to plan than when handed a fully specified brief

Weaknesses

  • Qwen's own label is an under-trained architecture preview, and low-bit quantization visibly costs quality — behaviour is a moving target rather than a fixed release
  • Very verbose in thinking mode — Artificial Analysis saw it burn roughly twice the median token count on its index, which quietly inflates output bills and stretches single tasks to many minutes
  • Ships a brand-new architecture and Qwen-XML tool calls — runtimes without the matching parser or kernels drop tool calls or decode several times slower than the active-parameter count implies

125B main model plus a separate 51B n-gram embedding table (176B on disk) and a 4B MTP head, 6B active per token; Qwen calls it an experimental preview of the Qwen4 architecture. GPQA Diamond 91.7 and HLE 35.9 are Qwen's own model-card figures (HLE judged by GPT-4o; the card does not say whether tools were used); Qwen also reports SWE-bench Pro 62.5, SWE-bench Multilingual 81.0, LiveCodeBench v6 91.9, CoWorkBench 73.9 and Toolathlon Verified 73.5, and the MMLU-Pro 73.23 in the card is for the pretrained Base checkpoint, not the released model. Terminal-Bench appears in the card's harness notes as the 2.0 version, with no figure found, so the 2.1 cell stays null. Arena has it at #24 overall in Agent Arena and 1617 (AutoEval) on Code Arena: WebDev, but no text-leaderboard Elo. Qwen states that "Qwen3.8-Flash is the official version based on Qwen3.8-Flash-Next with more production features, e.g., 1M context length by default, official built-in tools" — so pricing is Qwen's first-party rate for that hosted production version (1 RMB / 3 RMB, $0.16/$0.47; OpenRouter and Artificial Analysis list it at $0.15/$0.47), and costPerTask ($0.09) and speed (87.5 tok/s, 2.57 s to first answer token) are Artificial Analysis measurements taken on Alibaba's API without a disclosed effort tier (the model defaults to xhigh). Context is 262,144 native, extendable to about 1M via YaRN; max output is the 131,072 final-response cap (reasoning content may run to 262,144). Qwen's blog says it "plays the same role that Qwen3-Next played for Qwen3.5" and compares it to Qwen3.7-Plus — lineage language, not a replacement claim, so predecessorId is null.

For developers

API model strings

Not researched

Licence

Qwen Community License 1.0 · Restricted

Retirement

No retirement announced

Lineage

No recorded predecessor

News