Qwen 3.8 Flash
Qwen3.8 Flash is Alibaba's production-tier multimodal MoE reasoning model (125B total / ~6B active parameters) built on the open-weight Qwen3.8-Flash-Next architecture, exposing a 1M-token context by default and built-in tools at $0.15/$0.47 per 1M tokens — an aggressive price-performance play for agentic coding, long-context document/codebase analysis, and GUI automation available with Zero Data Retention on ARMES.
Empirical Evaluation Results
Architectural Profile & Capabilities
Qwen3.8 Flash is the hosted, production-serving variant of Qwen3.8-Flash-Next, released by Alibaba's Qwen team on August 26, 2026 (tech report committed to GitHub on August 24, 2026). The distinction matters architecturally: Flash-Next is the open-weight foundation model and is positioned by Qwen as a preview of the forthcoming Qwen4 generation, while Qwen3.8 Flash is the QwenCloud/OpenRouter production SKU that layers on 1M-token context by default, native tool schemas, adaptive tool use, and production safety filtering.\n\nARCHITECTURE. It is emphatically not a dense transformer and not a diffusion LLM — it is a sparse Mixture-of-Experts autoregressive decoder. Published specifications describe ~125B main parameters with only ~6B activated per token via routing, 48 layers, 512 total experts with 10 routed experts plus 1 shared expert active per token. Two large auxiliary banks sit outside the routed core: roughly 51B parameters of n-gram embedding tables and an ~4B multi-token-prediction (MTP) head, bringing total on-disk weights to roughly 180B. The layer design is a hybrid: Gated DeltaNet (GDN) linear-attention-style blocks combined with Qwen Sparse Attention (QSA), gated residual connections, and YaRN-style positional extension. QSA is a custom sparse long-context attention scheme rather than a plain MLA or GQA implementation; no vendor document brands it as either. Qwen reports QSA kernels delivering up to 7.6–8.6x prefill and ~4.9x decode throughput at 1M-token context versus the prior Qwen3.7-Plus implementation, and claims training compute on the order of one-ninth of Qwen3.7-Plus. The MTP head is explicitly a throughput play (speculative multi-token emission), which is why the effective serve speed materially exceeds what ~180B of stored weights would suggest.\n\nCONTEXT AND MULTIMODALITY. Native training context is 262,144 tokens; the 1M-token window is a YaRN-style extension rather than a natively trained window — an important distinction for retrieval degradation in the far tail. Modality coverage is text + image (chart, document, screenshot, video-frame), consistent with the OpenRouter 'image' modality flag. Independent vision evaluation (Roboflow Vision Evals) places it at ~68.8% average across six task families — strong on OCR (~88.7%) and structured data extraction (~84.5%), mid-pack on identification (~84.4%), and notably weak on object detection (~53.9%), counting (~56.8%), and visual reasoning (~54.1%). It scores 95.7 on MathVision, indicating that chart/diagram mathematics is a genuinely strong suit even where fine-grained spatial grounding is not.\n\nBENCHMARKS. Headline numbers come from the Qwen3.8-Flash-Next model card: GPQA Diamond 91.7 (above Claude Opus 4.6 Max's 91.3 in Qwen's own table), LiveCodeBench v6 91.9, SWE-bench Pro 62.5, SWE-bench Multilingual 81.0, DeepSWE 1.1 58.7, HLE 35.9, Toolathlon Verified 73.5, CoWorkBench 73.9 (vs 68.2 for Opus 4.6 Max), JobBench 55.7 (vs 36.6), AndroidWorld 84.5. Notably, Qwen publishes SWE-bench Pro rather than SWE-bench Verified — there is no citable vendor SWE-bench Verified, MATH-500, AIME, or MMMU figure for this SKU, and cross-lab comparisons should be treated accordingly. Independent aggregation gives Qwen3.8-Flash-Next ~56 on the Artificial Analysis Intelligence Index v4.1.1 at roughly 86 tokens/s output, and #19 on the LLM-Stats composite — i.e. frontier-adjacent reasoning at roughly 1/15th the per-token price of Western frontier tiers.\n\nOPERATIONAL PROFILE. Pricing is $0.15/1M input and $0.47/1M output (Novita-served on OpenRouter; QwenCloud lists ~$0.16/$0.47), with cached prompt prefixes at ~$0.016/1M — an ~89% discount that makes repeated long-context agentic loops economically viable. Measured output speed clusters around 54–90 tokens/s depending on provider, payload length and reasoning-token budget; the reasoning trace is billed as output and can inflate real-world cost 3–5x over the visible answer. Because it is a MoE with custom QSA/GDN kernels, self-hosting requires recent vLLM/SGLang builds with sparse-attention kernel support, and aggressive quantization can erode the QSA acceleration advantage.\n\nPOSITIONING. Against the September 2026 frontier (GPT-6 Astra, Claude Opus 5 / Sonnet 5, Gemini 3.8 Flash, DeepSeek V4 Pro, Grok 4.6, Kimi K3), Qwen 3.8 Flash is not a leaderboard-topper — Claude Opus 5 sits near 96% on SWE-bench Verified and Anthropic's preview models exceed 94% on GPQA Diamond. Its value proposition is instead a ~6B-active serving footprint that lands within a few points of last-generation frontier models on agentic coding and tool-use, with a 1M window and built-in tools at flash-tier pricing.
~125B total parameters with ~6B activated per token; 48 layers; 512 experts with 10 routed + 1 shared expert active per token; ~51B n-gram embedding bank plus ~4B multi-token-prediction head (~180B stored weights). Hybrid attention: Gated DeltaNet blocks interleaved with custom Qwen Sparse Attention (QSA) optimized for long context, plus gated residual connections and YaRN-style positional extension. Context: 262,144 tokens native, extended to 1,000,000 tokens (1M exposed by default in the production SKU). Multimodal text+image input via a vision encoder. Test-time compute: configurable reasoning/thinking tokens with an MTP head for multi-token speculative emission. Not MLA, not standard GQA, not a diffusion LLM.
Recommended Workloads & Primary Use Cases
- •1M context is YaRN-extended, not natively trained: fidelity is strongest below ~262K tokens and degrades in the far tail. Prefill time-to-first-token on near-1M payloads remains tens of seconds despite QSA's 7.6–8.6x prefill gains, so long-context calls should be batched or cached rather than made interactively.
- •Benchmark coverage is deliberately narrow: Qwen publishes SWE-bench Pro (62.5), not SWE-bench Verified, and there is no citable vendor MATH-500, AIME, MMMU or Terminal-Bench number for this SKU. Treat mixed-vendor leaderboards that quote a 'SWE-bench Verified' figure for Qwen3.8 Flash as unverified.
- •'Flash' branding is misleading relative to classic fast tiers: 6B active parameters plus a reasoning trace means perceived latency and billed output cost are dominated by thinking tokens (output billed at $0.47/1M). Disable or cap reasoning effort for latency-critical, non-analytical traffic.
- •Vision grounding is bimodal: excellent at text-rich and structured extraction (OCR ~88.7%, data extraction ~84.5%) but weak on fine localization (object detection ~53.9%, counting ~56.8%, visual reasoning ~54.1%). Do not deploy it for dense small-object detection or precise spatial reasoning without a dedicated vision model alongside it.
- •Deployment friction: the custom GDN + QSA hybrid and MoE routing require recent vLLM/SGLang builds with sparse-attention kernels; FP8/GGUF quantization and older inference stacks can silently lose the QSA speedup and inflate KV-cache memory at long context. Multi-turn agentic use also depends on interleaved-thinking and explicit tool-schema configuration, and it will over-call tools if schemas are loosely typed.
Calls to Qwen 3.8 Flash are routed through strict Zero Data Retention inference channels. Prompts and outputs are never stored, indexed, or monitored by Alibaba.
More Models from Alibaba
Compare other engines available in this laboratory.
Terminal-Bench 2.1: 86.6% (vendor) / 81.3% (AA independent), SWE-bench Pro: 67.7%, DeepSWE 1.1: 56.6%, PaperBench: 93.0% (#1), GPQA Diamond: 92.6%, FrontierSWE: 73.5%, OSWorld-Verified: 86.1%, OmniDocBench 1.5: 92.1%, IFBench: 82.8%, HLE: 43.6%, JobBench: 53.4% | Alibaba's most capable model and the first Qwen-Max released as open weights (GA Aug 3, 2026; open weights Aug 12, 2026).
Qwen3.8-27B is Alibaba's flagship open-weight dense vision-language model (Aug 2026): a 27B-parameter hybrid-attention multimodal transformer with thinking mode on by default, 262K native / 1M extended context, and frontier-tier coding, reasoning, and agentic scores at budget pricing ($0.42/$3 per 1M tokens).