Qwen3.8 is the August 2026 generation of large language models from Alibaba's Qwen Team, and this profile covers its two headline models: Qwen3.8-Max, the hosted flagship on Alibaba Cloud Model Studio built on the open Qwen3.8-2.4T-A95B weights, and Qwen3.8-27B, a dense open-weight vision-language model under Apache 2.0. It collects the published facts in one place: API identifiers, parameter counts, context windows, output limits, prices in every Model Studio region, licences, thinking controls, provider-reported benchmarks and lifecycle status. It is written for developers, architects, buyers and governance teams who need checked numbers rather than launch coverage. Every fact comes from Alibaba Cloud's documentation or the Qwen Team's Hugging Face model cards, listed under Standards and references. Where the provider publishes nothing on a point, the page says so.
Last checked: 2026-11-15.
This is an independent reference page from NHI Mgmt Group based on public provider documentation; it is not sponsored or endorsed by any provider. It uses Alibaba Cloud Model Studio's public documentation and the Qwen Team's Hugging Face model cards as of November 2026. Model details change, so confirm against the provider's current documentation.
Key takeaways
- Two models, two delivery routes. Qwen3.8-Max is an API model (qwen3.8-max, snapshot qwen3.8-max-0902) whose open base is Qwen3.8-2.4T-A95B; Qwen3.8-27B is an open checkpoint that Model Studio also serves as qwen3.8-27b.
- 2.4T total, 95B active versus 27B dense. The Max weights are a 512-expert Mixture-of-Experts model; the 27B is a dense model with a vision encoder.
- 262,144 tokens native, about 1M extended. Both open checkpoints list 262,144 tokens natively, extensible to 1,010,000 (2.4T-A95B) or 1,000,000 (27B); the hosted Max uses 1M by default.
- $2 / $6 for Max, $0.50 / $3 for 27B. Per million input and output tokens on the Singapore deployment; global-scope qwen3.8-max deployments in other regions list $1.65 and $4.951, with no separate thinking surcharge.
- Different licences. Qwen3.8-27B uses Apache 2.0; Qwen3.8-2.4T-A95B uses a licence named qwen3.8-max. Neither model appears on the Model Studio deprecation page as of 15 November 2026.
What is Qwen3.8?
Qwen3.8 is the fourth Qwen generation of 2026, after Qwen3.5, Qwen3.6 and Qwen3.7. The model card citations for the generation are dated August 2026, and the Max citation carries the title "Qwen3.8-Max: A New Bar for Coding and Cowork". Alibaba publishes Qwen in two forms at once: hosted API tiers (Max, Plus, Flash, Omni) on Model Studio, and open checkpoints on Hugging Face named by parameter count.
- Qwen3.8-Max: the Qwen3.8-2.4T-A95B card states that Qwen3.8-Max is the official version based on these weights, adding vision input, non-thinking support, a 1M context by default and official built-in tools. The open weights alone are text only and always think.
- Qwen3.8-27B: a causal language model with a vision encoder that takes text, images and video (including hour-scale video) and returns text. Thinking is on by default and can be turned off per request.
Both share the same hybrid layout of Gated DeltaNet (linear attention) and Gated Attention layers in a 3:1 ratio and were trained with multi-token prediction. For the full Qwen line-up, see Alibaba Qwen Models.
Key facts at a glance
As of 2026-11-15. Sources: Qwen3.8-2.4T-A95B and Qwen3.8-27B Hugging Face cards; Model Studio pricing page.
| Field | Value | Source |
|---|---|---|
| Provider | Alibaba (Qwen Team); hosted on Alibaba Cloud Model Studio | Model cards; pricing page |
| API ID | Max: qwen3.8-max, qwen3.8-max-0902 (open weights also hosted as qwen3.8-2.4t-a95b). 27B: qwen3.8-27b (Hugging Face Qwen/Qwen3.8-27B) | Pricing page; model cards |
| Release date | August 2026 (card citations); exact day not stated by Alibaba on the pages checked (2026-11-15) | Model cards |
| Status | Both active on Model Studio; both open weights | Pricing page; deprecation page; 27B card |
| Context window | Max: 1M by default (hosted); 262,144 native, up to 1,010,000 (weights). 27B: 262,144 native, up to 1,000,000 | Model cards |
| Max output | Both cards: up to 262,144 reasoning tokens and 131,072 final response tokens where the framework allows separate limits. Hosted API limit not stated by Alibaba on the pages checked (2026-11-15) | Model cards |
| Knowledge cutoff | Not stated by Alibaba on the pages checked (2026-11-15) | Model cards |
| Input modalities | Max: text (weights); vision added in the hosted version. 27B: text, image, video | Model cards |
| Output modalities | Text (both) | Model cards |
| Reasoning effort | xhigh (default), medium, low (both cards) | Model cards |
| Open weights | Yes (both) | Model cards |
| Licence | Max weights: qwen3.8-max licence. 27B: Apache 2.0 | Model cards |
| Parameters | Max: 2.4T total, 95B active. 27B: 27B (Hugging Face metadata shows 28B) | Model cards |
| Architecture | Max: 92 layers, 512 experts (10 routed plus 1 shared active). 27B: 64 dense layers | Model cards |
| Weight format | Max: BF16 safetensors | 2.4T-A95B card |
Release and version history
As of 2026-11-15. Alibaba Cloud does not publish a single release-notes timeline on the pages checked, so dates come from card citations and API snapshot names.
| Date | Event | Source |
|---|---|---|
| 2026-05-20 and 2026-06-08 | qwen3.7-max snapshots, the previous Max generation | Pricing page |
| August 2026 | Qwen3.8 open weights published: Qwen3.8-2.4T-A95B, Qwen3.8-27B and Qwen3.8-Flash-Next (card citations; the Max collection shows "Updated Aug 13") | Model cards |
| Date not stated | qwen3.8-max and the qwen3.8-max-0902 snapshot listed on Model Studio in six regions | Pricing page |
| Date not stated | qwen3.8-27b and qwen3.8-2.4t-a95b listed as hosted open-source models in Singapore; the 27B card still describes a Qwen Cloud version as "coming soon" | Pricing page; 27B card |
The undated name qwen3.8-max and the snapshot qwen3.8-max-0902 carry the same price in every region. Alibaba does not state on the pages checked which snapshot the undated name currently resolves to, so pin the snapshot ID where reproducible results matter.
How much does Qwen3.8 cost?
As of 2026-11-15. Price per million tokens, shown with a "$" sign on the pricing page (the page does not name a currency code). One tier covers 0 to 1M input tokens, and thinking and non-thinking output cost the same. Source: Model Studio pricing page.
| Item | Price | Notes | Source |
|---|---|---|---|
| qwen3.8-max, Singapore (International) | Input: $2; Output: $6 | 1M free tokens for 90 days; snapshot qwen3.8-max-0902 same price | Pricing page |
| qwen3.8-max, US (Virginia), US scope | Input: $2; Output: $6 | No free quota listed | Pricing page |
| qwen3.8-max, global scope (US Virginia, Germany Frankfurt, Hong Kong, Japan Tokyo) | Input: $1.65; Output: $4.951 | Marked as list prices; snapshot same price | Pricing page |
| qwen3.8-max, China (Beijing) | Input: $1.65; Output: $4.951 | Batch listed at a 50% discount for qwen3.8-max in this region | Pricing page |
| Context cache | Rates for qwen3.8-max not stated on the pricing page; it refers to the Context Cache page | General example on the page: explicit cache creation 125% and cache hits 10% of input | Pricing page |
| qwen3.8-27b, Singapore (International) | Input: $0.5; Output: $3 | Hosted open-source model; thinking and non-thinking; 1M free tokens; 0 to 1M tier | Pricing page |
| qwen3.8-2.4t-a95b, Singapore (International) | Input: $2; Output: $6 | Hosted open-source Max weights; 1M free tokens; 0 to 1M tier | Pricing page |
- Discount rules: batch and cache discounts cannot be combined, per the Qwen-Max section of the pricing page.
- Output billing: output prices cover chain of thought plus answer, so reasoning tokens are billed as output.
- Promotions: the page lists standard prices only; promotions appear in the Model Studio console.
Worked example at Singapore rates: a request with 100,000 input tokens and 10,000 output tokens (reasoning included) costs $0.20 for input plus $0.06 for output, $0.26 in total. See Model Pricing Explained.
Context window and output limits
The open checkpoints use a 262,144-token native window. Longer contexts rely on YaRN rope scaling, and the 27B card warns that static YaRN may hurt performance on shorter texts and suggests setting the scaling factor to the expected length (for example 2.0 for about 524,288 tokens). The hosted Qwen3.8-Max runs at 1M by default, so managed users do not configure scaling themselves.
| Limit | Value | Notes | Source |
|---|---|---|---|
| Native context (both checkpoints) | 262,144 tokens | No scaling required | Model cards |
| Extended context, Qwen3.8-2.4T-A95B | Up to 1,010,000 tokens | Self-hosted with scaling | 2.4T-A95B card |
| Extended context, Qwen3.8-27B | Up to 1,000,000 tokens | YaRN factor 4.0 in the card's example configuration | 27B card |
| Hosted Qwen3.8-Max context | 1M by default | Official version features | 2.4T-A95B card |
| Pricing tier | 0 to 1M input tokens | Single tier, no long-context surcharge | Pricing page |
| Reasoning content | Up to 262,144 tokens | Where frameworks allow separate limits | Model cards |
| Final response | Up to 131,072 tokens | Where frameworks allow separate limits | Model cards |
See Context Windows Explained.
Modalities and tools
- Qwen3.8-Max: the open weights are text only and multimodal input is not supported; the hosted version adds vision input and official built-in tools. The specific tool list is not stated on the pages checked.
- Qwen3.8-27B: text, image and video input. For higher frame-rate video sampling the card suggests raising longest_edge to 469,762,048 (about 224K video tokens); default sampling is 2 frames per second.
- Thinking on the Max weights: required for every interaction and cannot be disabled; responses begin with a reasoning block. The hosted Max adds non-thinking support.
- Thinking on the 27B: on by default; set enable_thinking to False per request for instruct mode.
- Reasoning effort: xhigh (default), medium and low on both cards. preserve_thinking is on by default, keeping reasoning from earlier turns in context.
- Recommended sampling: thinking mode temperature 1.0, top_p 0.95, top_k 20, presence_penalty 0.0; the 27B instruct mode uses temperature 0.7, top_p 0.80 and presence_penalty 1.5.
- Serving frameworks: official recipes for SGLang, vLLM and TokenSpeed; also Transformers, Docker Model Runner and community quantisations for llama.cpp, Ollama and LM Studio.
Built-in tools and agent harnesses give the model reach into systems; treat each connection as a credentialed integration. See the AI Coding Agents Security Guide.
Where is Qwen3.8 available?
As of 2026-11-15. Sources: Model Studio pricing page and model cards.
| Channel | Details | Source |
|---|---|---|
| Alibaba Cloud Model Studio | qwen3.8-max and qwen3.8-max-0902 in Singapore, US (Virginia), Germany (Frankfurt), Hong Kong (China), Japan (Tokyo) and China (Beijing); qwen3.8-27b and qwen3.8-2.4t-a95b in Singapore | Pricing page |
| Qwen Cloud | Named on the cards as the managed channel; a hosted Qwen3.8-27B is described as coming soon | Model cards |
| Hugging Face | Qwen/Qwen3.8-2.4T-A95B and Qwen/Qwen3.8-27B, with notebook links for Colab and Kaggle | Model cards |
| Local runtimes | Community quantisations for llama.cpp, Ollama and LM Studio (27B card; the 2.4T card also lists Jan) | Model cards |
| Other cloud marketplaces | Not stated by Alibaba on the pages checked (2026-11-15) | Pricing page; model cards |
The pricing page distinguishes deployment scopes: International (Singapore), Global (several regions) and US or Beijing scopes, so the region and scope chosen decide where calls are served. The hosted open-source rows (qwen3.8-27b, qwen3.8-2.4t-a95b) appear in the Singapore table; other regions for them are not stated on the portion of the page checked. Self-hosting the open checkpoints moves operation, patching and access control to the organisation; see the Self-Hosted LLM Security Guide.
Access restrictions
Neither model is described as restricted to invited or verified organisations on the pages checked. qwen3.8-max is listed with public prices in six regions, with a 1M-token free quota only in Singapore. Qwen3.8-27B and Qwen3.8-2.4T-A95B are downloadable from Hugging Face and also listed as hosted models; use is governed by each repository's licence file. Account-level rate limits for qwen3.8-max are not stated on the pages checked (2026-11-15).
Benchmarks (provider-reported, dated)
As of 2026-11-15. All figures are provider-reported by the Qwen Team on the Hugging Face cards (August 2026 generation). The 2.4T-A95B card labels its scores as Qwen3.8-Max, the hosted version, and reports no separate open-weight scores. Independent leaderboard results are not included because none were checked against their own published pages for this profile.
| Evaluation | Qwen3.8-Max | Qwen3.8-27B | Source and date |
|---|---|---|---|
| Terminal Bench 2.1 | 86.6 | 73.0 (Terminus) | Model cards, August 2026 |
| SWE-bench Pro | 67.7 | 61.7 | Model cards |
| DeepSWE 1.1 | 56.6 | 42.2 | Model cards |
| NL2Repo-Bench | 55.9 | 42.3 | Model cards |
| CoWorkBench | 74.8 | 70.7 | Model cards |
| JobBench | 53.4 | 33.4 | Model cards |
| Agents' Last Exam (pass / score) | 27.0 / 52.4 | 20.4 / 42.9 | Model cards |
| GPQA Diamond | 92.6 | 89.2 | Model cards |
| HLE (no tools) | 43.6 | 30.8 | Model cards |
| IFBench | 82.8 | 79.5 | Model cards |
| QwenSWEBench (internal) | 80.7 | 79.0 | Model cards |
- Max-only results: PaperBench 93.0, MRCR v2 256K (8-needle) 92.9, WideSearch 81.9, Toolathlon Verified 72.5, HealthBench 60.2, LongBench v2 66.3.
- 27B-only results (vision and agents): OSWorld-Verified 84.3, AndroidWorld 81.9, WebArena-Verified 64.8, OmniDocBench 1.5 91.1, LiveCodeBench v6 90.3.
- Harness notes on the 27B card: most coding and agent results use the Claude Code harness, HLE is judged by GPT-4o, and some vision scores are reported with and without a code interpreter.
Each card also lists comparator models chosen by the Qwen Team; those figures are the provider's selection and are not reproduced here. See How to Read LLM Benchmarks.
Model card and system card
The Qwen Team publishes a Hugging Face model card for each checkpoint. What the cards state, beyond the specifications above:
- Qwen3.8-2.4T-A95B card: architecture detail (hidden dimension 8192, 64 query and 4 key-value heads in the Gated Attention layers, 248,320-token padded vocabulary), the always-on thinking behaviour, recommended sampling and the list of hosted-only features.
- Qwen3.8-27B card: architecture (hidden dimension 5120, FFN intermediate dimension 17,408), YaRN configuration, video handling guidance, sampling for thinking and instruct modes, and a hosted version marked coming soon.
- Repetition guidance: both cards say presence_penalty can be raised between 0 and 2 to reduce endless repetition, at the risk of language mixing and slightly lower performance.
- System card: not stated by Alibaba on the pages checked (2026-11-15). Neither card includes a safety evaluation section.
See Model Cards and System Cards Explained.
Data handling
The Model Studio security and privacy page (last updated 11 October 2026) applies to hosted qwen3.8-max and qwen3.8-27b as to other Model Studio models; there is no model-specific data policy on the pages checked.
- Training use: the page states that Alibaba Cloud "will never use your data for model training".
- Retention: Model Studio stores data generated from model and application calls to comply with laws and regulations; the retention period is set in the related agreements, not on the page.
- Encryption: data transmitted during application building and model training is encrypted with AES-256. Encryption at rest is not addressed on the page.
- Assurance: Model Studio states SOC 2 compliance with an unqualified opinion, covering Security, Availability and Confidentiality.
- Self-hosted 27B or 2.4T-A95B: no data leaves the operator's environment through the model itself; retention and logging are the operator's own controls.
Model Studio API keys are non-human identities with direct spend and data exposure. See the LLMjacking Guide and the Alibaba Cloud Security Guide.
Deprecation status
As of 15 November 2026, the Model Studio deprecation page does not name any Qwen3.8 model. The page lists batches by date, with model IDs in linked announcements; the next scheduled batch is 19 January 2027. It states that accounts which called a model scheduled for retirement in the last three months are notified by email or internal message, and points to the Model Monitoring page in the Singapore or Beijing region to check usage.
| Model or item | Details | Source |
|---|---|---|
| qwen3.8-max, qwen3.8-max-0902, qwen3.8-27b, qwen3.8-2.4t-a95b | Not listed for deprecation (2026-11-15) | Model Studio deprecation page |
| Next scheduled batch | 2027-01-19 (models named in linked announcement) | Model Studio deprecation page |
| Qwen3.8-27B and Qwen3.8-2.4T-A95B weights | Open weights; downloads are not subject to the API deprecation schedule | Model cards |
See the LLM Deprecation and Retirement Tracker.
How Qwen3.8 differs from other models in its family
As of 2026-11-15. Sources: pricing page and Hugging Face cards. Models listed by size, largest first.
| Model | Parameters | Input modalities | Licence | Singapore price (in / out per 1M) | Positioning (provider wording) |
|---|---|---|---|---|---|
| Qwen3.8-Max (2.4T-A95B) | 2.4T total, 95B active | Text (weights); vision hosted | qwen3.8-max | $2 / $6 | "A New Bar for Coding and Cowork" (citation title) |
| Qwen3.8-Flash (Flash-Next) | 125B language model, 6B active (180B on Hugging Face metadata) | Text, image, video | qwen-community-1.0 | $0.15 / $0.47 (qwen3.8-flash) | "experimental preview of the architecture that will underpin Qwen4" |
| Qwen3.8-27B | 27B dense | Text, image, video | Apache 2.0 | $0.5 / $3 (qwen3.8-27b) | Dense vision-language model with thinking and instruct modes |
The practical differences are licence (only the 27B is Apache 2.0), hardware (a 27B dense model versus a 2.4T-parameter MoE), modality (the Max weights are text only) and thinking control (the Max weights cannot disable thinking; the 27B can). Qwen3.8-Omni-Flash, a separate audio and video model, is outside this profile. For cross-provider context see Open-Weight Frontier Models Compared and Mixture of Experts Explained.
Practitioner checklist
- Decide between hosted (qwen3.8-27b, qwen3.8-max) and self-hosted weights based on data location, licence and operating cost.
- Call qwen3.8-max-0902 rather than the undated name where reproducible behaviour matters, and record the snapshot in your AI inventory.
- Choose the Model Studio region and deployment scope deliberately; global scope and US scope carry different prices and serving locations.
- Budget for reasoning tokens: output prices include chain of thought, and the open Max weights cannot turn thinking off.
- Read the licence file before use: Apache 2.0 for Qwen3.8-27B, the qwen3.8-max licence for Qwen3.8-2.4T-A95B.
- For self-hosting beyond 262,144 tokens, set the YaRN factor to the expected length and test short-prompt quality before rollout.
- Re-run your own evaluations; the published scores are provider-reported and partly from internal or harness-specific benchmarks.
- Store Model Studio API keys in a secrets manager, separate them by environment and set spend alerts.
- Check the Model Studio deprecation page and Model Monitoring each month for Qwen3.8 entries.
Standards and references
- Qwen Team: Qwen3.8-2.4T-A95B model card (accessed 2026-11-15)
- Qwen Team: Qwen3.8-27B model card (accessed 2026-11-15)
- Qwen Team: Qwen3.8-Flash-Next model card (accessed 2026-11-15)
- Alibaba Cloud: Model Studio model pricing (accessed 2026-11-15)
- Alibaba Cloud: Model Studio models (accessed 2026-11-15)
- Alibaba Cloud: Model Studio model deprecation (accessed 2026-11-15)
- Alibaba Cloud: Model Studio security certifications and privacy (accessed 2026-11-15)
Related NHI Mgmt Group resources: Alibaba Qwen Models, Chinese vs US Frontier Models Compared, Open-Weight Model Security Guide, Frontier AI Models Compared.