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).
Qwen3.8-27B is an image-text-to-text causal language model with a vision encoder, released by Alibaba Qwen in August 2026 as the highest-tier open-weight VL reasoning model in the Qwen3.8 family. It is a DENSE model (27B active parameters, no MoE routing) built on the Qwen-Next / Qwen3.5-3.6 architectural foundation, with a notable hybrid attention layout: 64 layers arranged as 16 × (3 × (Gated DeltaNet → FFN) → 1 × (Gated Attention → FFN)). Only 16 of 64 layers run full (softmax) Gated Attention — configured GQA-style with 24 query heads / 4 KV heads at head-dim 256 — while the remaining 48 layers use Gated DeltaNet linear attention (recurrent-memory updates with separate QK/V head partitioning), which dramatically cuts KV-cache growth and decode FLOPs at long context. The model card reports native context of 262,144 tokens, extendable to 1,000,000 tokens (the figure listed on OpenRouter), though most official evals are run at 256K. Thinking mode is enabled by default with controllable reasoning_effort levels (xhigh/medium/low) and preserve_thinking support, indicating a reasoning-oriented post-training pipeline (the core config is essentially unchanged from Qwen3.6-27B, with gains attributed to alignment/RL optimization). Verified performance is exceptional for the size class: SWE-bench Pro 61.7 (Claude Code harness, temp 1.0/top_p 0.95, 256K ctx), LiveCodeBench v6 90.3, GPQA Diamond 89.2, Terminal-Bench 2.1 73.0, OSWorld-Verified 84.3, Agents' Last Exam 42.9 score (20.4 Pass@1), IFBench 79.5, and an Artificial Analysis Intelligence Index of 52 at max reasoning effort — positioning it between GPT-5.6-class frontier APIs and much larger open MoE models, at roughly 1/10th the input cost of Claude Opus-class APIs. Vision-side capability (MMMU ~81.7 per secondary reporting) makes it a strong multimodal agent backbone.
Architecture Type: Dense Autoregressive Transformer (hybrid linear + full attention) with vision encoder
27B dense parameters, 64 layers, hidden size 5,120. Hybrid attention: 16 × (3 × Gated DeltaNet linear-attention → FFN) → 1 × Gated Attention → FFN; only 16/64 layers use full attention (GQA: 24 query heads / 4 KV heads, head-dim 256; DeltaNet blocks use separate QK/V partitioning ~16/16). SwiGLU FFN, RMSNorm, causal LM with vision encoder and multimodal projector. Thinking mode (reasoning tokens) ON by default with reasoning_effort xhigh/medium/low; native 262,144-token context, extendable to 1,000,000.
Recommended Workloads & Primary Use Cases
Long-running software engineering agents (SWE-bench Pro 61.7, Terminal-Bench 73.0)
Multimodal document intelligence & OCR-heavy professional workflows
Complex scientific/technical reasoning with tunable thinking effort (GPQA Diamond 89.2)
Long-context research synthesis over 256K-1M token corpora
Cost-efficient coding copilots and code review pipelines (LiveCodeBench v6 90.3)
GUI/OS automation and computer-use agents (OSWorld-Verified 84.3)
Operational Caveats & Boundaries
•1M-token context is an extension claim: native training context is 262,144 tokens and official evals run at 256K — expect degraded retrieval precision and higher TTFT beyond ~256K, especially on needle-in-haystack tasks.
•Thinking mode is ON by default and emits reasoning tokens billed at the $3/1M output rate; without explicit reasoning_effort control, agentic loops can burn output budget and inflate latency. Use low/medium effort for high-volume extraction tasks.
•Hybrid linear attention (48/64 Gated DeltaNet layers) trades long-range recall fidelity for speed — very long-context exact-recall and cross-document reference tasks are weaker than full-attention models of similar class.
•SWE-bench Pro 61.7 is Qwen's own eval using the Claude Code harness (temp 1.0, top_p 0.95) and is not directly comparable to SWE-bench Verified leaderboard numbers; treat agentic scores as harness-dependent.
•Vision input resolution caps and tiling rules are not fully documented in public sources; high-res images expand token counts substantially, which interacts with (and can consume) the 1M context budget.
ARMES Zero Data Retention
Calls to Qwen-3.8-27b are routed through strict Zero Data Retention inference channels. Prompts and outputs are never stored, indexed, or monitored by Alibaba.
