Qwen3.6
v20260416Alibaba
Alibaba's Apache-2.0 open-weight Qwen3.6 family (Apr 2026): Qwen3.6-35B-A3B MoE (35B total / 3B active, 2026-04-16) and Qwen3.6-27B dense (2026-04-22). Hybrid Gated DeltaNet + Gated Attention with thinking mode and Thinking Preservation; 262K native context (~1M via YaRN); text, image, and video input across 201 languages. The 27B beats the 397B-A17B Qwen3.5 flagship on agentic coding (77.2% SWE-bench Verified). The newer Qwen3.7-Max/Plus frontier remains API-only.
Trust Vector Analysis
Dimension Breakdown
🚀Performance & Reliability+
Remarkable capability density: the 27B dense model beats Alibaba's own 397B-A17B Qwen3.5 flagship on agentic coding (77.2% SWE-bench Verified), and the 35B-A3B delivers near-flagship reasoning with 3B active parameters. Main caveats: vendor benchmarks run above independent evaluations, and default-on thinking mode is token-hungry.
Vendor agentic-coding benchmarks corroborated by independent analysis; independent harness runs tend to land below vendor peaks
Mathematical and scientific reasoning benchmarks from official model cards; vendor-reported, pending broader third-party replication
Knowledge and multimodal benchmark review across 201-language coverage; smaller models trail the 397B flagship on knowledge breadth
Repeated-prompt and multi-turn agent testing, supplemented by community reports since April 2026
Independent throughput measurements on reference hardware plus architecture analysis
Qualitative assessment; per-provider p95 distributions not yet broadly published
Official specification from model cards
Hosted-platform availability plus redundancy across third-party hosts and self-hosting
🛡️Security+
Inherits the Qwen family's solid multilingual guardrails, but the April 2026 launch means little independent red-teaming exists yet, and image/video input widens the attack surface. Open weights shift responsibility to deployers who fine-tune.
Testing against OWASP LLM01 patterns including image/video-borne injection; limited third-party data given April 2026 launch
Adversarial prompt testing; assessment accounts for open-weight modifiability and launch recency
Analysis of hosted-platform policies plus the self-hosting option for full data isolation
Safety testing across harmful content categories and multiple languages on default weights
Review of API security features on the first-party hosted platform
🔒Privacy & Compliance+
Same posture as Qwen3.5: Alibaba's first-party API is China-jurisdiction (Singapore region available), which concerns Western regulated buyers; Apache-2.0 self-hosting or Western third-party hosting fully avoids that. No HIPAA/FedRAMP for the model service. The small footprints (27B dense fits a single 24GB GPU in 4-bit) make compliant self-hosting unusually practical.
Review of hosting regions and licensing; China-jurisdiction caveat applies to Alibaba's first-party API, not self-hosted or Western-hosted deployments
Analysis of hosted-platform data usage terms
Review of hosted-platform retention policies; retention is deployment-dependent for open-weight models
Review of data protection capabilities and customer responsibilities
Verification of infrastructure certifications versus model-service-level compliance for Western regulated markets
Review of data handling across first-party API, third-party hosts, and self-hosting
👁️Trust & Transparency+
Strong open documentation and unusually inspectable reasoning via default-on thinking with Thinking Preservation. Typical Qwen gaps remain: limited training-data detail, topic-avoidance on politically sensitive subjects, and — given the recent launch — sparse independent factuality data.
Evaluation of reasoning transparency and trace accessibility
Early factual QA and grounding observations; limited independent data given April 2026 launch
Evaluation on bias benchmarks across languages and politically sensitive topic probes
Qualitative assessment of confidence expression in outputs
Review of model card and technical documentation completeness
Review of public disclosures about training data
Analysis of built-in safety mechanisms in default weights
⚙️Operational Excellence+
Rides the mature Qwen ecosystem: Apache 2.0 with patent grant, day-one vLLM/SGLang/KTransformers/transformers support, and hosted options from Alibaba Cloud ($0.60/$3.60 per 1M for 27B) to cheaper third-party hosts. Only two sizes released so far — the family ladder is thinner than Qwen3.5's 0.8B-397B range — and the newest family frontier (Qwen3.7-Max/Plus) is API-only.
Review of API design, consistency, and feature completeness
Review of SDK and inference-framework support
Review of release cadence and weight-availability guarantees
Review of monitoring tools across deployment options
Assessment of support tiers, documentation, and community responsiveness
Analysis of derivative models, third-party hosting, and tooling integrations; scored slightly conservative given the ~3-month age
Review of licensing terms and restrictions
- +27B dense model beats the 397B-A17B Qwen3.5 flagship on agentic coding: 77.2% SWE-bench Verified, 59.3% Terminal-Bench 2.0
- +Exceptional efficiency: 27B runs in ~18GB VRAM (4-bit) on a single 24GB GPU; 35B-A3B activates only 3B parameters
- +Apache 2.0 with patent grant across both released models
- +Multimodal input (text, image, video) across 201 languages and dialects
- +262K native context, extensible to ~1M tokens via YaRN
- +Thinking Preservation keeps reasoning traces across turns, improving iterative agent workflows
- +Day-one vLLM/SGLang/KTransformers/transformers support within the largest open-model ecosystem
- !First-party Alibaba Cloud hosting is China-jurisdiction (Singapore region available); no HIPAA/FedRAMP path for the model service — self-hosting or Western hosts avoid this
- !Vendor benchmarks notably exceed independent evaluations; still below closed frontier models on the most complex tasks
- !Token-hungry and relatively slow in default thinking mode (~56 tok/s for the 27B self-hosted)
- !Only two sizes released (27B dense, 35B-A3B) versus Qwen3.5's full 0.8B-397B ladder; family frontier Qwen3.7-Max/Plus is API-only
- !Topic-avoidance on politically sensitive subjects in default weights
- !Training-data composition disclosed only at a high level
- !Launched April 2026: limited independent red-teaming and factuality studies so far
Use Case Ratings
code generation
77.2% SWE-bench Verified from a 27B dense model that self-hosts on a single 24GB GPU in 4-bit — exceptional agentic-coding economics; beats the 397B-A17B Qwen3.5 flagship.
customer support
201-language coverage with a 3B-active MoE variant that serves high-volume tiers very cheaply; thinking mode should be toggled off for latency.
content creation
Strong multilingual content with image and video understanding for visually grounded writing.
data analysis
Native image/video input handles charts and documents; 262K context (~1M via YaRN) covers large datasets at small-model cost.
research assistant
Multimodal document understanding, long context, and preserved reasoning traces suit iterative research workflows.
legal compliance
First-party hosting is China-jurisdiction; viable for regulated legal work only via self-hosting or certified Western hosts.
healthcare
No HIPAA path on first-party hosting; the small footprint makes self-hosted deployment in compliant infrastructure practical.
financial analysis
Strong quantitative reasoning (AIME 2026 94.1%) with chart/table understanding; data-residency planning required for regulated workloads.
education
201 languages, multimodal input, visible reasoning traces, and single-GPU deployability make it excellent for global and budget education deployments.
creative writing
Capable multilingual creative output with visual grounding; prose distinctiveness behind dedicated creative leaders.