Model comparison
Qwen3.8-2.4T-A95B vs Qwen3.8-27B
Two Qwen tiers compared on the figures that decide which one a workload actually needs.
Catalog record checked August 14, 2026; individual provider fields may change.
Qwen
Qwen3.8-2.4T-A95B
Frontier · Open weights
Qwen
Qwen3.8-27B
Balanced · Open weights
| Specification | Qwen3.8-2.4T-A95B | Qwen3.8-27B |
|---|---|---|
| Provider | Qwen | Qwen |
| Tier | Frontier | Balanced |
| Context window | 262K | 262K |
| Max output | 131K | 131K |
| Input / 1M tokens | Not verified | Not verified |
| Output / 1M tokens | Not verified | Not verified |
| Weights | Open | Open |
| Parameters | 2.4T total / 95B active (MoE) | 27B dense |
| Reasoning levels | low, medium, xhigh | low, medium, xhigh |
| Modalities | text | text, image, video |
| Released | August 12, 2026 | August 14, 2026 |
Prices are USD per million tokens at standard rates, excluding batch and caching discounts. Bold indicates the better figure where one is objectively better. Values we could not confirm from the provider are shown as “Not verified” rather than estimated.
Qwen3.8-2.4T-A95B
Official Hugging Face checkpoint for Qwen3.8 open weights — a text base model, not the hosted Qwen3.8-Max API.
Best for
- Self-hosting the Qwen3.8 flagship
- vLLM / SGLang serving
- Evaluating the open Max-class weights
Watch out
Native context is 262,144 tokens (YaRN extends toward ~1,010,000). Hugging Face best-practice caps: 262,144 reasoning tokens and 131,072 final-response tokens. Thinking cannot be disabled. Qwen3.8-Max License: MaaS / AI Work Assistant businesses over $50M trailing 12-month revenue need a separate commercial licence. Self-hosting still needs datacentre-class memory. Do not treat this checkpoint as the multimodal 1M-context API.
Qwen3.8-27B
Apache 2.0 Qwen3.8 dense VLM for local and self-hosted work — native 262K context, image and video input, thinking on by default but can be turned off. Distinct from hosted Qwen3.8-Max.
Best for
- Local multimodal agents
- Self-hosting a dense 27B VLM
- Apache 2.0 deployments
Watch out
The Hugging Face repo was created 2026-08-05; this row uses the 2026-08-14 card revision. Hosted 1M-context API is documented as coming soon. Vendor SWE-bench Pro 61.7 and DeepSWE 42.2 used a Claude Code harness, not swebench.com or the public Datacurve DeepSWE board. YaRN can extend context toward 1M.
Common questions
Qwen3.8-2.4T-A95B vs Qwen3.8-27B
Answered from the verified figures on this page rather than general guidance.
Which has the larger context window, Qwen3.8-2.4T-A95B or Qwen3.8-27B?
Both accept about 262K tokens of context, so document length will not decide between them.
Do Qwen3.8-2.4T-A95B and Qwen3.8-27B support the same reasoning levels?
Yes — both accept the same effort settings: "low", "medium", "xhigh". Higher effort costs more and takes longer, so start low and raise it only where output quality actually improves.
Should I use Qwen3.8-2.4T-A95B or Qwen3.8-27B?
Qwen3.8-2.4T-A95B is the frontier tier and Qwen3.8-27B the balanced tier. The useful question is whether your hardest task actually fails on the cheaper one — most production volume such as classification, extraction and summarisation does not.
Next step
Choosing between them
Tier and workload decide this more reliably than a leaderboard position does.
If both sit in the same tier, the decision usually comes down to context window and output price rather than headline capability — output tokens dominate real bills.
If one is a step up within the same provider, the useful question is whether your hardest task actually fails on the cheaper tier. Most production volume — classification, extraction, summarization — does not.
Choosing a harness rather than a model — Cursor, Copilot, Claude Code, Muse Code, or Lovable? Compare agentic harnesses · Latest releases.