Qwen3.8-Flash-Nextreview, pricing and limits

Released by Alibaba's Qwen team on August 26, 2026, Qwen3.8-Flash-Next extends to a 1-million-token context via YaRN and trained at roughly one-ninth the compute of its Qwen3.7-Plus predecessor. It suits self-hosted coding and long-context agents more than teams needing a managed API today.

Qwen3.8-Flash-Next is Alibaba's open-weight preview of the Qwen4 architecture, released August 26, 2026, scoring 91.7 on GPQA Diamond and 62.5 on SWE-bench Pro. It activates just 6 billion of roughly 180 billion parameters per token and natively handles a 262,144-token context.

Provider: Alibaba Cloud · Family: Qwen3.8

More about Alibaba Cloud on HokAI

Context window: 262,144 tokens · Max output: 131,072

Input modalities: text, image, video, tool-calls · Output: text, tool-calls

About Qwen3.8-Flash-Next

Qwen3.8-Flash-Next is a foundation model from the Qwen Team at Alibaba Cloud, released August 26, 2026, its first and so far only entry. It is a multimodal Mixture-of-Experts model with a 125 billion parameter main network, a 51 billion parameter N-gram embedding table, and a 4 billion parameter multi-token prediction module, roughly 180 billion parameters combined, of which only 6 billion activate per token. Alibaba frames it as an early, open-weight preview of the architecture underpinning the coming Qwen4 flagship, sitting in the Qwen3.8 generation alongside Qwen3.8-Max and Qwen3.8-27B, benchmarked against predecessor Qwen3.7-Plus in its own technical report. That report shows the model scoring 91.7 on GPQA Diamond, 62.5 on SWE-bench Pro (55.8 for Qwen3.7-Plus, 53.4 for Claude Opus 4.6 Max), 81.0 on SWE-bench Multilingual, 58.7 on DeepSWE 1.1 (54.4 for DeepSeek-V4-Flash-0731), and 91.9 on LiveCodeBench v6. The one benchmark where Alibaba concedes the lead is Humanity's Last Exam, 35.9 versus Claude's 40.0, and it also trails on NL2Repo-Bench, 48.1 to 54.2. These are vendor-run evaluations on Alibaba's own tooling, with some SWE-bench Pro tasks corrected by the team and GPT-4o used as grader for part of the suite. The model natively handles 262,144 tokens of context, extends to 1,000,000 via YaRN RoPE scaling, and outputs up to 131,072 tokens. Its Qwen Sparse Attention component gives up to 7.6x faster prefill and 4.9x faster decode at full context versus Qwen3.7-Plus, with prefix-cache throughput 8.6x higher at a 90% hit rate. Inputs cover text, image, and video: it scores 95.7 on MathVision with a code interpreter and 88.5 on RealWorldQA, and tool use is evaluated via CoWorkBench (73.9), AndroidWorld (84.5), and Toolathlon Verified (73.5), with OpenAI-compatible function calling and structured JSON outputs. There is no native audio input or output. Qwen3.8-Flash-Next has no first-party hosted API price; it ships only as open weights on Hugging Face and ModelScope. Alibaba separately announced a differently named model, Qwen3.8-Flash (no "Next"), for the Qwen Cloud API at $0.16 per million input tokens and $0.47 per million output tokens, but that had not gone live at release and is not the same model. Running it means self-hosting: the FP8 checkpoint is 172.78 GiB and BF16 is 335.28 GiB, needing a multi-GPU node at minimum tensor-parallel degree 2 on GB300-class hardware, TP4 for production. Supported stacks include vLLM, SGLang, and llama.cpp for GGUF quants, with fine-tuning via Unsloth, ms-swift, and LLaMA-Factory. The 51 billion N-gram embedding parameters sit in system RAM rather than GPU memory, changing the usual VRAM-only sizing math. No system card, refusal-rate benchmark, or training data cutoff date is published for this release. It supports the standard Qwen post-training toolchain (SFT, DPO, GRPO) for teams fine-tuning their own alignment, and uses the Muon optimizer alongside AdamW on select weight categories, cutting training compute to roughly one-ninth of Qwen3.7-Plus's. Thinking mode is on by default, with a reasoning_effort switch (xhigh, medium, low) and preserved-thinking that carries the reasoning trace across a conversation, meant to help agentic sessions keep context. Weights use the Qwen Community License 1.0, not Apache 2.0: commercial use, modification, and redistribution are allowed, but products with over 100 million monthly users or $20 million in monthly revenue must display the model name, and a hosted model-as-a-service API or coding/office assistant built on it needs a separate license. It fits multi-GPU teams wanting near-frontier coding and agent performance at a fraction of typical training cost, or anyone evaluating Qwen4's architecture early. It is a poor fit for teams needing a managed API today, single-GPU deployments, or products requiring a system card before shipping; DeepSeek-V4-Flash-0731 and Qwen3.7-Plus are the closest hosted alternatives named in its own benchmarking.

Pricing

This model carries no vendor per-token rate: it is distributed purely as downloadable weights, not sold through a metered API. A separately named production model, Qwen3.8-Flash (no "-Next"), is instead the one Alibaba assigned a Qwen Cloud rate for hosted use, but that product had not gone live as of this release and is a distinct model, not this one. See the pricing FAQ below for the exact hosted figures.

Key Features

• GDN + QSA Hybrid Attention: Combines Gated DeltaNet with Qwen Sparse Attention across 48 layers, cutting long-context latency and lifting prefix-cache throughput 8.6x at a 90% cache hit rate.
• 1M-Token Extended Context: Extends via YaRN RoPE scaling to a 1,000,000-token window from its native size, with a 131,072-token maximum output.
• 6B Active Parameters per Token: A large MoE backbone plus a separate N-gram embedding table and a multi-token prediction module, but only 6B parameters activate on any single forward pass.
• Vision and Video Input: Reads images and video alongside text, posting an 88.5 score on the RealWorldQA benchmark and strong results on chart and document reasoning tasks.
• Tool-Use and Agent Benchmarks: Supports OpenAI-compatible function calling and structured JSON outputs, scoring 84.5 on AndroidWorld and 73.5 on Toolathlon Verified for agentic tool use.

Pros

• Beats Qwen3.7-Plus and DeepSeek-V4-Flash on multiple coding benchmarks (62.5 vs 55.8 on SWE-bench Pro, 58.7 vs 54.4 on DeepSWE 1.1) while activating only 6B parameters per token.
• 1M-token extended context with QSA attention delivering up to 7.6x faster prefill at full context than Qwen3.7-Plus.
• Open weights on Hugging Face and ModelScope with same-day FP8 and GGUF-friendly releases, so it can be self-hosted and inspected immediately.

Cons

• No first-party hosted API: running it means self-hosting a 172-plus GiB checkpoint on a multi-GPU node, even in the smaller FP8 format.
• No published training data cutoff date or system card at release, unlike most frontier model launches.
• Trails Claude Opus 4.6 Max on Humanity's Last Exam (35.9 vs 40.0) and DeepSeek-V4-Flash-0731 on NL2Repo-Bench (48.1 vs 54.2).

Benchmarks

• ifbench: 81.3
• lvbench: 76.6
• jobbench: 55.7
• coworkbench: 73.9
• deepswe 1 1: 58.7
• realworldqa: 88.5
• androidworld: 84.5
• gpqa diamond: 91.7
• nl2repo bench: 48.1
• swe bench pro: 62.5
• livecodebench v6: 91.9
• charxiv reasoning: 90.6
• mathvision with ci: 95.7
• humanitys last exam: 35.9
• toolathlon verified: 73.5
• swe bench multilingual: 81

Visit Qwen3.8-Flash-Next Official Page