Introducing HIPAA-compliant Qwen3.8-27B

The open-weight releases worth reading about used to divide into two camps: trillion-parameter MoE flagships nobody can self-host, and tiny local models nobody should trust with real work. Every so often a release lands in the middle - genuinely capable, genuinely deployable - and those are the ones worth a healthcare team's attention.

Today, we are bridging that gap. Qwen3.8-27B, Alibaba's new dense 27 billion parameter vision-language model, is now available on our HIPAA-compliant API platform under a signed BAA and strict zero data retention - with no proprietary vendor lock-in. See the full specs →

What is Qwen3.8-27B?

Released to Hugging Face around August 13-14, 2026 (repo created August 5, hosted API "coming soon" on Qwen Cloud), Qwen3.8-27B is a dense model - a single 27B parameter stack, no Mixture-of-Experts routing - under the Apache 2.0 license. That licensing choice is the cleanest in the Qwen3.8 family and the reason we think this release matters commercially, not just technically.

The architecture carries the Qwen3.8 hybrid design into a dense shape: a 64-layer stack mixing Gated DeltaNet (linear attention) and Gated Attention in a repeating pattern, with a 248,320 padded token embedding, multi-token prediction trained in, and a 262,144 token context window natively, extensible to 1,000,000 tokens. It is a native vision-language model - a vision encoder handles images and videos alongside text from the start, not as a bolt-on adapter. On the API, thinking is on by default with a non-thinking mode available, and reasoning_effort exposes xhigh (default), medium, and low.

The Apache 2.0 license and mixed attention architecture make it genuinely deployable outside a data center: community-available FP8 and GGUF builds run it on a single GPU, and the smallest quantizations fit in about 8GB of memory. That combination - license, size, and capability - is what we think earns it a spot in the HIPAA-compliant conversation.

Performance profile: what it's good at vs. what it's not

What Qwen3.8-27B excels at

• Agentic coding, genuinely Opus-class in shape. On its model card (Terminus harness at 256K context, temperature 1.0, top_p 0.95), it scores 73.0 on Terminal-Bench 2.1 and 61.7 on SWE-bench Pro, where the Qwen3.8 card lists Opus 4.6-class numbers at 78.2 and 53.4 respectively - it beats that Opus figure on SWE-bench Pro outright, a result that would have been implausible from a 27B dense model two releases ago.
• Deployment economics. At 27B dense with no expert routing, the serving cost profile sits far below anything in the trillion-parameter class, and the Apache 2.0 license means the deployment path has no revenue ceiling to read carefully.
• Native vision and video. On Qwen's card, it scores 84.3 on OSWorld-Verified computer use (versus 72.7 for the Opus figure in the same row), 64.8 on WebArena-Verified browser use, and 38.6 on SWE-MM multimodal software engineering. On RealWorldQA it hits 85.9 and on LVBench long-video understanding 72.4.
• Reasoning-quality headroom for the size. GPQA Diamond at 89.2 and HLE at 30.8 are, again per the model card, solidly above the previous Qwen3.6-27B's 87.8 and 24.0 - a real capability step, not a benchmark refresh.
• A real intermediate tier. For teams tiering models by task sensitivity, this is the natural mid-tier pick - a smaller, faster, cheaper sibling of the Qwen3.8 architecture rather than a downscale of something that works differently.

Limitations to keep in mind

• 27B dense is not a frontier write-off, but it is a tier below. On Terminal-Bench 3.0 and 4.0 - benchmarks published after its build date - the Qwen3.8 family's own frontier models and Claude-class checkpoints score meaningfully higher than the 30.0 and 31.2 this model posts. Long-horizon agentic repairs, multi-day coding tasks, and open-ended research synthesis still belong to the bigger tiers.
• Vendor-reported benchmarks on a new harness. Qwen's card tables mix self-run and third-party-borrowed baselines, and agentic benchmarks are famously harness-sensitive. We would want third-party reruns before treating any single number as settled.
• No native medical-imaging or document parity. Being a vision model does not make it a radiology tool or a DICOM reader. It reads charts, screenshots, documents, and interface states - not diagnostics - and we make no clinical-accuracy claim for the vision path.
• Context claims vs. context practice. "262K native, extensible to 1M" is the model's architecture, not what every serving stack honors. On self-hosted FP8 and quantized GGUF builds, real usable context at good throughput is lower; on a managed API, the hosted tier's configured limit is the operative number.
• Timing. The hosted version with production features ("1M context by default, official built-in tools") was still marked "coming soon" on the model card at the time of writing - building against that version, rather than the open weights, is a different deployment plan.

The benchmarks: how Qwen3.8-27B compares

Benchmarks are from the official Qwen3.8-27B model card on Hugging Face (vendor-reported, Terminus/Claude Code harness at 256K context, temperature 1.0, top_p 0.95), with the "Opus4.6 Max" column as Qwen reports it.

The multimodal agentic rows are what separate this release from a generic 27B text model: on OSWorld-Verified computer use and WebArena-Verified browser use, it posts numbers in the same territory as the frontier reference column, not merely ahead of its own predecessor.

Why run Qwen3.8-27B via our HIPAA-compliant API?

• Enterprise-grade privacy. Available under our HIPAA-compliant design-partner program with a signed BAA, end-to-end data encryption, and strict zero-data-retention on prompts and outputs. Your patient data is never used for training. General availability is coming soon.
• Uncompromised performance. We handle serving, context-window scaling, and the vision encoder pathway - you get a vision-capable, agentic, deployment-friendly model without operating the inference hardware yourself.
• Transparent pricing. Under Apache 2.0 with 27B dense parameters and no expert-routing overhead, the per-token economics are among the cheapest of any model in our catalog at $0.80 per 1M input, $0.40 per 1M cached input, and $4.00 per 1M output - a fraction of what an Opus-class API costs on Azure OpenAI or AWS Bedrock.

Ready to deploy Apache 2.0 agentic vision under your BAA? Join the waitlist now.