📊 See Qwen3.8-Max rankings

📱 Get the AI Rank app

Qwen 3.8 comes in two sizes that matter for most people. Qwen3.8-Max is Alibaba's 2.4-trillion-parameter flagship, a mixture-of-experts model with about 95B parameters active per token. It costs $2 per million input tokens and $6 per million output tokens on QwenCloud's international endpoint. Qwen3.8-27B is a dense vision-language model under Apache-2.0 that fits on a single high-end GPU once quantized. On Alibaba's own benchmark tables, Max leads 27B on every shared coding and agent test, by margins of 1.7 to 20 points.

Max launched through the API on August 2, 2026. Its open weights followed on August 12 as Qwen3.8-2.4T-A95B, the first time a Max-class Qwen model has been released openly, and the 27B checkpoint arrived on August 14.

Qwen3.8-27B vs Qwen3.8-Max Specs

Qwen3.8-27B Qwen3.8-Max (open weights: Qwen3.8-2.4T-A95B)
Architecture Dense, 64 layers, hybrid Gated DeltaNet + gated attention MoE, 92 layers, 512 routed experts (10 active) + 1 shared
Parameters 27B 2.4T total, about 95B active
Modalities Text, image, video Open weights: text only
Context 262,144 native, up to 1M with YaRN 262,144 native, up to 1M; API tier covers up to 1M input
License Apache-2.0 Qwen3.8-Max License (custom)
Hosted API QwenCloud listing "coming soon" qwen3.8-max, snapshot qwen3.8-max-0902

Both models think by default. Reasoning depth is set per request with reasoning_effort (xhigh by default, plus medium and low), and preserve_thinking keeps earlier reasoning in multi-turn agent loops. The QwenCloud snapshot qwen3.8-max-0902 (dated September 2, 2026) adds native vision on the hosted side, but the open 2.4T checkpoint is a text-only causal LM.

Qwen 3.8 Benchmarks: 27B Against Max

All scores below are self-reported by the Qwen team, taken from the Qwen3.8-27B model card and the Qwen3.8-Max launch blog. Where a benchmark appears in both tables, the setup is the same unless noted.

Benchmark Qwen3.8-27B Qwen3.8-Max Gap
SWE-bench Pro 61.7 67.7 6.0
DeepSWE 1.1 42.2 56.6 14.4
NL2Repo-Bench 42.3 55.9 13.6
QwenSWEBench (in-house) 79.0 80.7 1.7
CoWorkBench (in-house) 70.7 74.8 4.1
JobBench 33.4 53.4 20.0
Agents' Last Exam (Pass@1 / Score) 20.4 / 42.9 27.0 / 52.4 6.6 / 9.5
IFBench 79.5 82.8 3.3
GPQA Diamond 89.2 92.6 3.4
HLE 30.8 43.6 12.8

Terminal Bench 2.1 is left out of the table on purpose. The 27B card reports 73.0 using the Terminus harness, while the Max blog reports 86.6 using Claude Code (avg@10, five-hour timeout), so the two numbers come from different setups. In the Max blog's own comparison, Max trails Claude Fable 5 on SWE-bench Pro (67.7 vs 80.0) and GPT-5.6 Sol (max) on Terminal Bench 2.1 (86.6 vs 88.8), but it posts the top IFBench score in that table at 82.8.

Human-preference data tells a similar story. On the Arena text leaderboard published October 15, 2026, Qwen3.8-Max sits at #24 overall with a rating of 1482 (1477 to 1487, 27,766 votes), and Qwen3.8-27B sits at #100 with 1438 (1433 to 1443, 23,294 votes).

→ Track it on AI Rank: Qwen3.8-27B — Arena rating, rank and 90-day history

Qwen 3.8 API Pricing

QwenCloud and Alibaba Cloud Model Studio list qwen3.8-max at the same rate in thinking and non-thinking modes, with reasoning tokens billed as output. Prices are in USD per million tokens:

Endpoint Input Output Implicit cache input Explicit cache create / read
Singapore (International scope) $2.00 $6.00 $0.25 $2.50 / $0.17
Frankfurt, Virginia, Tokyo, Hong Kong (Global scope) and Beijing $1.65 $4.951 $0.206 $2.063 / $0.137

Beijing also offers batch-file calls at half price ($0.825 input, $2.475 output). A free quota of one million tokens applies in Singapore only and lasts 90 days from account activation or model release, whichever is later. Maximum output per request is 131K tokens.

Alibaba does not offer a hosted Qwen3.8-27B yet. Its model card says the QwenCloud version, with 1M context by default, is coming soon. Until then, third-party hosts fill the gap. OpenRouter's catalog listed qwen/qwen3.8-27b at $0.04 input and $1.35 output per million tokens on October 22, 2026, and qwen/qwen3.8-2.4t-a95b at $2 and $6. Those are OpenRouter's prices, not Alibaba's.

→ Live usage: See where both models rank this week in the OpenRouter model usage ranking on AI Rank (tokens, Oct 10–16, 2026 window).

In that window, Qwen3.8-27B was #34 with 387 billion tokens (down 23% week over week), and the qwen3.8-max-20260902 snapshot was #50 with 262 billion.

Running Qwen 3.8 Locally

The 27B is the practical local model. Official BF16 weights are at Qwen/Qwen3.8-27B, with an official FP8 build at Qwen/Qwen3.8-27B-FP8; both are also on ModelScope. Qwen points to SGLang, vLLM and TokenSpeed for serving. The vLLM recipe lists these layouts:

Hardware Checkpoint Notes from the vLLM recipe
1 Blackwell GPU NVFP4 (Inferact/Qwen3.8-27B-NVFP4) Weights about 24.6 GiB; 262K context
1× RTX 5090 NVFP4 Needs --enforce-eager; context cut to 32K
2× RTX 5090 NVFP4, tensor parallel 2 Full 262K context
4 GPUs (one GB300 tray) Qwen/Qwen3.8-27B-FP8 Largest KV cache

A single-GPU launch looks like this:

vllm serve Inferact/Qwen3.8-27B-NVFP4 \
  --tensor-parallel-size 1 \
  --max-model-len 262144 \
  --kv-cache-dtype fp8 \
  --reasoning-parser qwen3 \
  --enable-auto-tool-choice --tool-call-parser qwen3_coder

Qwen's recommended sampling for thinking mode is temperature=1.0, top_p=0.95, top_k=20. For non-thinking mode, it recommends temperature=0.7, top_p=0.8 and presence_penalty=1.5.

Max is a datacenter job. vLLM's recipe runs the NVFP4 build on one 8-GPU B300 node (tensor parallel 8). FP8 needs about 2,325 GiB, which means 16 GPUs across two nodes. The license is also different. The Qwen3.8-Max License allows commercial use, deployment and fine-tuning, with two conditions. Products with more than 100 million monthly users or $20 million in monthly revenue must show the model name in the interface. Companies running a model-as-a-service or AI coding and office assistant business with more than $50 million in revenue over 12 months need a separate license from Qwen.

Where Qwen3.6 Fits

Qwen3.6 is the spring generation that Qwen3.8 replaces. Qwen3.6-35B-A3B came out on April 16, 2026 and Qwen3.6-27B on April 22, both as open weights. On the Qwen3.8-27B card, the older 27B scores 53.5 on SWE-bench Pro and 63.4 on Terminal Bench 2.1, against 61.7 and 73.0 for the new one. Qwen3.6-35B-A3B still draws traffic: it was #22 on OpenRouter for the October 10 to 16 week with 836 billion tokens.

Compare before you choose: pricing changes and rankings move daily — check the current model rankings on AI Rank before you commit.

Or 📱 get the AI Rank app to follow them on your phone.

Sources

Prepared with automated editorial assistance and checked on October 22, 2026 against Qwen's blog, model cards, license and GitHub repository, QwenCloud and Alibaba Cloud pricing pages, vLLM recipes, the OpenRouter catalog and dated AI Rank data. We did not run our own benchmarks for this article.