All tools / Coding & Development / Qwen 3.8
Model & API reviewQwen 3.8 Review 2026: Models, Benchmarks, API Pricing & Comparisons
Qwen 3.8 is Alibaba Qwen's August–September 2026 flagship model family—combining a 2.4T-parameter hosted and open-weight Max tier, an Apache-2.0 27B vision-language dense model, and a 6B-active Flash-Next preview—built on a hybrid Gated DeltaNet and sparse MoE architecture for agentic coding, long-context work, and lower API cost.
Updated November 15, 2026 · Information is based on public sources, not hands-on testing; confirm current endpoints, weights licenses, and Model Studio tariffs with Qwen.
Direct answer: should you pick Qwen 3.8?
Choose Qwen 3.8 when you want near-frontier coding-agent and long-horizon reasoning quality at mid-tier API prices ($0.15–$2.00 per 1M input tokens) or need an Apache-2.0 27B multimodal model you can self-host locally. On vendor-reported and public coding evaluations, Qwen3.8-Max reaches 86.6 on Terminal-Bench 2.1, 67.7 on SWE-bench Pro, and 92.6 on GPQA Diamond, while its open-weight 2.4T-A95B and 27B releases give engineering teams deployment options that closed frontier labs do not ship.
Skip or qualify Qwen 3.8 first if your procurement policy prohibits non-Apache custom licenses on the flagship weights (qwen3.8-max on 2.4T-A95B and qwen-community-1.0 on Flash-Next), if you require full multimodal vision inside the open-weight 2.4T download (the 2.4T open checkpoint is text-only), or if your team is already standardized on Claude Code with Anthropic Opus/Sonnet credits.
Qwen 3.8 model family breakdown
Rather than shipping a single checkpoint, Qwen split the August–September 2026 rollout across five distinct tiers. Confusing the hosted API variants with the open-weight Hugging Face repositories is the most common evaluation mistake.
| Variant | Architecture & Size | Modalities & Context | License / Access | Primary Role |
|---|---|---|---|---|
| Qwen3.8-Max / Max-0902 | 2.4T total / 95B active MoE (512 experts, 92 layers) | Text, image, video input; 262K native, 1M hosted (131K output) | Hosted API (Qwen Cloud / Model Studio / OpenRouter) | Flagship coding agents, web front-end generation, 1M-token research |
| Qwen3.8-2.4T-A95B | 2.4T total / 95B active MoE (same core backbone) | Text-only (no vision encoder); 262,144 native context | Open weights (qwen3.8-max license on HF & ModelScope) |
Cluster-hosted private reasoning, repository agents, enterprise distillation |
| Qwen3.8-27B | 27B dense language (28B total w/ vision encoder, 64 layers) | Native text, image & video; 262K native, 1M via YaRN/hosted | Open weights (Apache-2.0) + hosted API | Single-GPU/Apple Silicon local coding, VLM document parsing, permissive commercial SaaS |
| Qwen3.8-Flash / Flash-Next | 125B MoE + 51B N-gram embeddings (6B active; Qwen4 preview) | Vision-language (QSA sparse attention + Gated Residual); 262K–1M | qwen-community-1.0 open weights (Flash-Next) & hosted Flash API |
High-throughput agents, fast subagents, low-cost routing |
| Qwen3.8-Omni-Flash & LiveTranslate | End-to-end omnimodal / speech-text streaming | Audio, speech, video, 60-language simultaneous translation | Hosted API (Sept 18, 2026 release) | Real-time voice agents and live multilingual captioning |
Architecturally, Qwen3.8-2.4T-A95B and Qwen3.8-27B combine 3× Gated DeltaNet linear attention layers with 1× Gated Attention per block, plus Multi-Token Prediction (MTP) training. Meanwhile, Qwen3.8-Flash-Next previews Qwen's next-generation stack—replacing standard Gated Attention with Qwen Sparse Attention (QSA) at the micro-block level, adding Gated Residual streams, and scaling token capacity via N-gram embeddings so only 6B parameters activate per token.
Qwen 3.8 benchmarks vs. frontier & open alternatives
The figures below combine Qwen's official August–September 2026 model-card evaluations (tested under the Claude Code, mini-SWE-agent, and Codex harnesses) and public leaderboard standings. Because benchmarks vary by harness timeout and scaffolding, treat cross-lab numbers as directional and verify on your own repository tasks.
| Benchmark | Qwen3.8-Max | Qwen3.8-27B | Claude Opus 4.8 | GPT-5.6 Sol (max) |
|---|---|---|---|---|
| Terminal-Bench 2.1 (avg@10) | 86.6 (Claude Code) | 73.0 | 84.6 (Terminus 2) | 88.8 (Codex) |
| SWE-bench Pro | 67.7 | 61.7 | 69.2 | 64.6 |
| FrontierSWE | 73.5 | — | 70.0 | — |
| PaperBench | 93.0 | — | 80.3 | 90.5 |
| GPQA Diamond | 92.6 | 89.2 | 92.0 | 94.1 |
| IFBench (Instruction Following) | 82.8 | — | 62.2 | 72.7 |
| MRCR v2 256K (8-needle) | 92.9 | — | 83.2 | 93.8 |
| HealthBench / Legal (PLawBench) | 60.2 / 73.2 | — | 52.4 / 69.6 | 55.3 / 72.3 |
Beyond the initial August launch, the September 2026 Qwen3.8-Max-0902 snapshot climbed to #1 on the Code Arena WebDev leaderboard with a 1,691 Elo score, and Qwen3.8-Max reached 85.6% on Vals AI's SWE-bench tracking. Just as notable is Qwen3.8-27B: scoring 73.0 on Terminal-Bench 2.1 and 89.2 on GPQA Diamond puts a single-GPU open-weight model ahead of many full-size 2025 frontier models.
API pricing, thinking controls & local hardware requirements
Qwen 3.8 prices sit aggressively below closed US frontier tiers while keeping explicit controls for reasoning token spend. Thinking mode is enabled by default and can be toggled off per request, tuned via reasoning_effort, or preserved across multi-turn agent loops via preserve_thinking.
| Model / Competitor | Input (per 1M) | Output (per 1M) | Cache Read / Notes | Local Serving Option |
|---|---|---|---|---|
| Qwen3.8-Max (Model Studio Intl) | $2.00 ($1.65 Global/CN) | $6.00 ($4.951 Global/CN) | $0.25 cached input; 1M context | Open 2.4T-A95B needs multi-node H200/B200 cluster |
| Qwen3.8-27B | $0.50 Intl ($0.035 OpenRouter) | $3.00 Intl ($1.50 OpenRouter) | Apache-2.0 weights; native VLM | 4-bit GGUF fits 16–24 GB VRAM GPU or 24–32 GB Apple Silicon |
| Qwen3.8-Flash / Flash-Next | ~$0.15 | ~$0.47 | 6B active MoE; 1M hosted | Open Flash-Next weights (~6B active footprint) |
| Kimi K3.1 Pro (comparison) | $2.00 | $10.00 | $0.20 cache hit; 1M context | See our Kimi K3.1 Pro profile |
Worked API cost example: Sending a 120,000-token repository context (with 100,000 cached prefix tokens and 20,000 uncached tokens) plus generating 4,000 output tokens on Qwen3.8-Max international pricing costs (100,000 × $0.25 + 20,000 × $2.00 + 4,000 × $6.00) / 1,000,000 = $0.089 per agent turn—substantially less when routing routine subtasks to Qwen3.8-Flash or self-hosting Qwen3.8-27B via vLLM or SGLang (see our DeepInfer vs vLLM vs SGLang guide).
Best for, when to skip & top alternatives
Best for
- Agentic coding harnesses: multi-file bug fixes, repository navigation, and frontend React/UI scaffolding where Terminal-Bench 2.1 and SWE-bench Pro reliability matter.
- Cost-tiered routing: pairing
Qwen3.8-Maxfor planning/hard reviews withQwen3.8-FlashorQwen3.8-27Bfor bulk execution. - Local & private VLMs: teams needing an Apache-2.0 27B dense model with native image/video understanding and 262K context on a single workstation GPU.
When to skip (choose another tool when)
- Strict Apache/MIT legal mandate on flagship weights:
Qwen3.8-2.4T-A95Buses the customqwen3.8-maxlicense andFlash-Nextusesqwen-community-1.0; only27Bis Apache-2.0. - Turnkey IDE UX without setup: developers wanting zero configuration inside an editor may prefer Cursor or Claude Code.
- Single-vendor US compliance: organizations bound to US-only providers may choose Claude, ChatGPT, or compare DeepSeek.
Evaluation checklist before adopting Qwen 3.8
- Confirm the exact model ID and checkpoint date (
Qwen3.8-Max-0902vs. earlier August snapshots, or open2.4T-A95Bvs.27B) because license terms and vision support differ across them. - Inspect
reasoning_effortand non-thinking toggles on your target prompt set; reasoning tokens count toward billed output tokens even when hidden from final UI output. - Verify regional endpoint terms (Alibaba Cloud Model Studio International vs. China vs. OpenRouter) for latency, data residency, and prompt-caching rates.
- Run your own repository test suite rather than relying solely on published Terminal-Bench 2.1 or SWE-bench Pro scores, especially if you use custom tool-calling schemas.
- For self-hosting
Qwen3.8-27BorFlash-Next, validate vLLM/SGLang kernel support for Gated DeltaNet and QSA layers before sizing GPU memory.
Frequently asked questions
What models are in the Qwen 3.8 family?
The lineup includes hosted Qwen3.8-Max (and Max-0902), open-weight text-only Qwen3.8-2.4T-A95B, Apache-2.0 vision-language Qwen3.8-27B, high-speed Qwen3.8-Flash / Flash-Next (125B MoE + 51B N-gram embeddings, 6B active), plus Qwen3.8-Omni-Flash and Qwen3.8-LiveTranslate.
Is Qwen 3.8 open weight or proprietary?
It is mixed by tier. Qwen3.8-27B uses the permissive Apache-2.0 license, Qwen3.8-2.4T-A95B ships under the custom qwen3.8-max license, and Qwen3.8-Flash-Next uses qwen-community-1.0. Hosted Qwen3.8-Max-0902, Omni-Flash, and LiveTranslate are API-only.
How much does the Qwen 3.8 API cost compared with frontier competitors?
Model Studio international pricing sits near $2.00 input / $6.00 output per 1M tokens for Qwen3.8-Max ($1.65/$4.951 on Global/China endpoints, $0.25 cached read), ~$0.50/$3.00 for Qwen3.8-27B (lower on third-party hosts like OpenRouter), and ~$0.15/$0.47 for Qwen3.8-Flash—under half the output cost of Kimi K3.1 Pro ($10/1M).
Can Qwen3.8-27B run locally on consumer hardware?
Yes. Quantized 4-bit GGUF builds of Qwen3.8-27B run comfortably on 16–24 GB VRAM workstation GPUs or 24–32 GB Apple Silicon Macs while retaining native image/video input and a 262,144-token context window.
How does Qwen 3.8 compare against Claude, DeepSeek, and Kimi K3.1 Pro?
On Qwen's August 2026 evaluations, Qwen3.8-Max reaches 86.6 on Terminal-Bench 2.1, 67.7 on SWE-bench Pro, and 93.0 on PaperBench—competitive with Claude Opus 4.8 and GPT-5.6 Sol—while offering both a lower API price and downloadable open-weight tiers. Compare workflows in our AI coding assistants guide.
What is the difference between Qwen3.8-2.4T-A95B and hosted Qwen3.8-Max?
Qwen3.8-2.4T-A95B is the downloadable text-only open-weight checkpoint capped at 262K native context under the qwen3.8-max license. Hosted Qwen3.8-Max includes multimodal vision/video encoders, 1M context by default, built-in tools, and the newer Max-0902 checkpoint.