THE AI RANKINGS

Alibaba

Qwen 3.6

Provider
Alibaba
Status
available
Context
262,144 tok
SWE-bench
77.2%
Price
$0.6 / $3.6 /MTok

Qwen 3.6 is Alibaba’s April 2026 open-weight generation — released under the permissive Apache 2.0 licence, so the weights are free to download, self-host and use commercially (Qwen on GitHub). It ships as two open models: the dense Qwen3.6-27B flagship (22 April) and the efficient Qwen3.6-35B-A3B mixture-of-experts with 3B active parameters (16 April). Both use a hybrid attention design, support a 256K-token context (extensible to ~1M via YaRN) and accept text, image and video input.

Its standout result is that the compact 27B model outperforms the far larger Qwen 3.5 (a 397B MoE) on agentic coding, scoring 77.2% on SWE-bench Verified (Hugging Face). It is the open counterpart to Alibaba’s closed flagship Qwen3.7-Max, and sits in the Chinese open-weight cluster with DeepSeek V4, MiniMax M3 and Kimi.

Quick specs

ProviderAlibaba (Qwen)
ReleasedApril 2026 (35B-A3B 16 Apr, 27B 22 Apr)
StatusAvailable (open weights)
Open modelsQwen3.6-27B (dense), Qwen3.6-35B-A3B (MoE)
ArchitectureHybrid Gated DeltaNet + gated attention
Context window262,144 tokens (≈1M via YaRN)
ModalitiesText + image + video in, text out; multilingual
LicenceApache 2.0 (open weights, self-hostable)
Self-host priceFree (≈18GB VRAM for the 27B in 4-bit)
SWE-bench Verified77.2% (27B, vendor)
Best forSelf-hosted multilingual/multimodal coding and agentic work
LimitationsToken-hungry/slow per independent testing; below the closed frontier

VIEW QWEN 3.6 →

What Qwen 3.6 is

“Qwen 3.6” is a generation, not a single model. The open release centres on two Apache-2.0 models (GitHub):

Both use a hybrid attention architecture combining Gated DeltaNet (a linear-attention mechanism) with gated self-attention layers, support a 262,144-token native context extensible to about 1,010,000 tokens via YaRN scaling, and accept text, image and video input (Hugging Face). They run a hybrid thinking mode on by default — reasoning can be toggled off — and add “Thinking Preservation,” which keeps reasoning traces across a conversation. Multilingual coverage carries over from the Qwen 3.x lineage (reported at 201 languages and dialects).

Above the open tier sits the closed Qwen3.6-Max-Preview flagship — Alibaba’s first closed-weight flagship — and, a month later, the closed Qwen3.7-Max. This page covers the open Qwen 3.6 models; the Max tier is proprietary and API-only.

Benchmark performance

Vendor figures are from Alibaba’s model cards; Artificial Analysis provides a more conservative independent read.

BenchmarkQwen3.6-27BNotes
SWE-bench Verified77.2Beats the 397B Qwen 3.5 on agentic coding (HF)
GPQA Diamond87.8Vendor-reported
AIME 202694.1Vendor-reported
LiveCodeBench v683.9Vendor-reported
MMLU-Pro86.2Vendor-reported
Artificial Analysis Intelligence Index46 → 3746 at launch (open-weights leader <150B); 37 on a later revised index (Artificial Analysis)

The picture is a strong, efficient open-weight model — punching above its size on coding and reasoning vendor benchmarks — tempered by independent testing, where Artificial Analysis rates it accurate but token-hungry and relatively slow (~56 tokens/sec). The smaller 35B-A3B scores a few points lower (SWE-bench Verified 73.4, GPQA 86.0) but is far cheaper to run. The closed frontier (GPT-5.5, Claude Opus 4.8) still leads on the hardest tasks. See best AI for coding for cross-model standings.

Pricing and access

The open weights are free to download and self-host under Apache 2.0 — the 27B runs in about 18GB of VRAM in 4-bit, on a single 24GB GPU (buildfastwithai). Weights are on Hugging Face and ModelScope as Qwen/Qwen3.6-27B and Qwen/Qwen3.6-35B-A3B, with vLLM, SGLang and llama.cpp support.

Hosted pricing varies by provider: the 27B is roughly $0.60 input / $3.60 output per million tokens via Alibaba’s API (Artificial Analysis) and about $0.29 / $3.17 on OpenRouter; the smaller 35B-A3B is about $0.14 / $1.00 on OpenRouter. It is also served through Alibaba Cloud Model Studio and the consumer Qwen app.

How Qwen 3.6 compares

Known limitations

Below the closed frontier on the hardest reasoning and agentic tasks. Token-hungry and slow — Artificial Analysis flags high output-token counts and ~56 tok/s, which raises real-world cost and latency. China-hosted API — the hosted route carries data-residency and content-control considerations (self-hosting the open weights avoids this). Benchmark spread — vendor figures run notably higher than the independent intelligence index, and the AA score itself shifted between index versions; prefer standardized leaderboards where a decision rides on the number.

FAQ

What is Qwen 3.6?

Qwen 3.6 is Alibaba’s April 2026 open-weight generation — primarily a dense 27B model and a 35B-A3B mixture-of-experts, both under Apache 2.0, with a 256K context (extensible to ~1M via YaRN) and text, image and video input.

Is Qwen 3.6 open source?

Yes — the open Qwen 3.6 models are released as open weights under Apache 2.0, free to download, self-host, fine-tune and use commercially, on Hugging Face and ModelScope. (The separate Qwen3.6-Max-Preview and Qwen3.7-Max flagships are closed.)

How much does Qwen 3.6 cost?

The open weights are free to self-host. Hosted, the 27B is about $0.60 input / $3.60 output per million tokens via Alibaba’s API and ~$0.29 / $3.17 on OpenRouter; the 35B-A3B is about $0.14 / $1.00 on OpenRouter.

How good is Qwen 3.6 at coding?

Strong for its size — the 27B scores 77.2% on SWE-bench Verified (vendor), beating the much larger Qwen 3.5 — though the closed frontier models still lead, and independent testing rates it accurate but slow.

What is the difference between Qwen 3.6 and Qwen3.7-Max?

Qwen 3.6 is the open generation (Apache 2.0, self-hostable). Qwen3.7-Max is the closed, API-only flagship above it — more capable, with a 1M-token context, but not downloadable.


Last verified 19 June 2026. Architecture, Apache-2.0 licence, context window and the headline SWE-bench figure are confirmed via Qwen’s GitHub and Hugging Face model cards; pricing via Artificial Analysis and OpenRouter. Benchmark numbers are largely vendor-reported and run higher than the independent Artificial Analysis index (shown with both the launch and revised figures). Confirm against current leaderboards before relying on them.