Alibaba's Qwen Team shipped three Qwen generations between April and August 2026, Qwen3.6, Qwen3.7 and Qwen3.8, and each one changed API names, prices, open checkpoints and thinking controls. This page compares the three generations side by side: API IDs and snapshots, release dates, open weights and licences, context windows, output guidance, Model Studio prices, documented behaviour changes, provider-reported benchmarks and lifecycle status. It is written for developers moving workloads between Qwen versions, buyers estimating cost and architects deciding between a hosted tier and a self-hosted checkpoint. Generations are listed newest first, an objective order that implies no ranking. Every fact comes from Alibaba Cloud Model Studio documentation or the Qwen Team's Hugging Face model cards, cited under Standards and references.
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
- Max got cheaper in Qwen3.8. qwen3.8-max lists $2 input and $6 output per million tokens on the Singapore deployment, against $2.5 and $7.5 for qwen3.7-max, both on a single 0 to 1M tier.
- Qwen3.7 has no open checkpoint on the pages checked. Qwen3.6 and Qwen3.8 both publish open weights on Hugging Face; the Qwen3.7 models checked (Max, Plus, Flash) appear only as hosted APIs.
- The 27B dense model kept its shape. Qwen3.8-27B and Qwen3.6-27B share the same 64-layer hybrid layout and dimensions; the Qwen3.8 card reports higher scores on every shared benchmark row, including SWE-bench Pro 61.7 against 53.5.
- Qwen3.8 adds reasoning effort levels. The Qwen3.8 cards document reasoning_effort values xhigh (default), medium and low, and preserve_thinking on by default; the Qwen3.6 cards describe preserve_thinking as an option and no effort setting.
- No Qwen3.6, 3.7 or 3.8 retirement is announced. Model Studio's deprecation page names no model IDs in its batches; stable models get at least 3 months' notice and preview models at least 14 days.
What is Qwen?
Qwen is the large language model family from Alibaba's Qwen Team. Alibaba publishes Qwen in two forms at once: hosted API tiers on Alibaba Cloud Model Studio (Max, Plus and Flash, each with dated snapshot IDs) and open checkpoints on Hugging Face named by parameter count, such as Qwen3.6-35B-A3B, where A3B means about 3 billion active parameters. A version number such as 3.8 marks a generation, and a generation can contain hosted tiers, open checkpoints or both.
The three generations differ in what they publish. Qwen3.6 has hosted Plus and Flash tiers plus two Apache 2.0 checkpoints. Qwen3.7 has hosted Max, Plus and Flash tiers. Qwen3.8 has hosted Max and Flash tiers plus two open checkpoints, one of which, Qwen3.8-2.4T-A95B, is the base of the hosted Qwen3.8-Max. For the full Qwen line-up, see the Alibaba Qwen Models hub; for the newest generation in depth, see the Qwen3.8 (Max and 27B) profile.
Every model at a glance
As of 2026-11-15. Source: Model Studio pricing page and Qwen Hugging Face model cards. Undated API names resolve to the snapshot shown where the pricing page states it.
| Model | API ID or repository | Generation | Form | Status |
|---|---|---|---|---|
| Qwen3.8-Max | qwen3.8-max, qwen3.8-max-0902 | 3.8 | Hosted; based on Qwen3.8-2.4T-A95B | Active |
| Qwen3.8-Flash | qwen3.8-flash | 3.8 | Hosted | Active |
| Qwen3.8-2.4T-A95B | Qwen/Qwen3.8-2.4T-A95B | 3.8 | Open weights | Published |
| Qwen3.8-27B | Qwen/Qwen3.8-27B | 3.8 | Open weights; hosted version "coming soon" on Qwen Cloud per card | Published |
| Qwen3.7-Max | qwen3.7-max (resolves to qwen3.7-max-2026-05-20), qwen3.7-max-2026-06-08 | 3.7 | Hosted | Active |
| Qwen3.7-Max preview | qwen3.7-max-preview (resolves to qwen3.7-max-2026-05-17) | 3.7 | Hosted, thinking only | Preview |
| Qwen3.7-Plus | qwen3.7-plus (resolves to qwen3.7-plus-2026-05-26) | 3.7 | Hosted | Active |
| Qwen3.7-Flash | qwen3.7-flash (resolves to qwen3.7-flash-2026-07-15) | 3.7 | Hosted | Active |
| Qwen3.6-Plus | qwen3.6-plus (resolves to qwen3.6-plus-2026-04-02) | 3.6 | Hosted | Listed on pricing page |
| Qwen3.6-Flash | qwen3.6-flash (resolves to qwen3.6-flash-2026-04-16) | 3.6 | Hosted | Listed on pricing page |
| Qwen3.6-35B-A3B | Qwen/Qwen3.6-35B-A3B | 3.6 | Open weights | Published |
| Qwen3.6-27B | Qwen/Qwen3.6-27B; hosted as qwen3.6-27b per card | 3.6 | Open weights and hosted | Published |
Licences: Qwen3.6-35B-A3B, Qwen3.6-27B and Qwen3.8-27B use Apache 2.0; Qwen3.8-2.4T-A95B lists a licence named qwen3.8-max. The hosted tiers are API only. Knowledge cutoff: not stated by Alibaba on the pages checked (2026-11-15) for any model here.
When was each model released?
As of 2026-11-15. Alibaba Cloud does not publish a single release-notes timeline on the pages checked, so dates are API snapshot dates from the pricing page and release months from model card citations.
| Date | Event | API ID or repository |
|---|---|---|
| 2026-04-02 | qwen3.6-plus snapshot | qwen3.6-plus-2026-04-02 |
| 2026-04-16 | qwen3.6-flash snapshot | qwen3.6-flash-2026-04-16 |
| April 2026 | Qwen3.6 open weights; card titles "Agentic Coding Power, Now Open to All" (35B-A3B) and "Flagship-Level Coding in a 27B Dense Model" (27B) | Qwen/Qwen3.6-35B-A3B, Qwen/Qwen3.6-27B |
| 2026-05-17 | qwen3.7-max preview snapshot, thinking only | qwen3.7-max-2026-05-17 |
| 2026-05-20 | qwen3.7-max snapshot | qwen3.7-max-2026-05-20 |
| 2026-05-26 | qwen3.7-plus snapshot | qwen3.7-plus-2026-05-26 |
| 2026-06-08 | Second qwen3.7-max snapshot | qwen3.7-max-2026-06-08 |
| 2026-07-15 | qwen3.7-flash snapshot | qwen3.7-flash-2026-07-15 |
| August 2026 | Qwen3.8 open weights; citation title "Qwen3.8-Max: A New Bar for Coding and Cowork" | Qwen/Qwen3.8-2.4T-A95B, Qwen/Qwen3.8-27B |
| Not stated | qwen3.8-max with snapshot qwen3.8-max-0902, and qwen3.8-flash, listed on Model Studio | qwen3.8-max, qwen3.8-flash |
Snapshot dates are the dates encoded in the IDs, not announcement dates. The pricing page does not state which snapshot the undated qwen3.8-max resolves to, so pin qwen3.8-max-0902 where reproducible results matter. See LLM Naming and Versioning Explained.
How much does each model cost?
As of 2026-11-15. Price per million tokens on the Singapore (International) deployment, shown with a "$" sign on the pricing page. Thinking and non-thinking output cost the same for these models. Source: Model Studio pricing page.
| Model | Input tier | Input | Output | Free quota | Notes stated by provider |
|---|---|---|---|---|---|
| qwen3.8-max (and -0902) | 0 to 1M | $2 | $6 | 1M tokens | Context caching discount |
| qwen3.8-flash | 0 to 1M | $0.15 | $0.47 | 1M tokens | Context caching discount |
| qwen3.7-max (and dated snapshots) | 0 to 1M | $2.5 | $7.5 | 1M tokens (undated name only) | Context caching discount |
| qwen3.7-max-preview | 0 to 1M | $2.5 | $7.5 | None | Thinking only; no discount labels |
| qwen3.7-plus | 0 to 256K / 256K to 1M | $0.4 / $1.2 | $1.6 / $4.8 | 1M tokens | "Limited-time 20% off" label on the undated name |
| qwen3.7-flash | 0 to 32K / 32K to 256K / 256K to 1M | $0.030 / $0.100 / $0.200 | $0.130 / $0.400 / $0.800 | 1M tokens | 50% batch discount; context caching discount |
| qwen3.6-plus | 0 to 256K / 256K to 1M | $0.5 / $2 | $3 / $6 | None | No discount labels |
| qwen3.6-flash | 0 to 256K / 256K to 1M | $0.25 / $1 | $1.5 / $4 | None | 50% batch discount; context caching discount |
- Free quota: valid for 90 days from Model Studio activation, model release or application approval, whichever is later, and listed only on the Singapore deployment.
- Batch and cache: the Max section states batch calls cost 50% of real-time prices where supported, and the two discounts cannot be combined.
- Standard prices only: the page states that promotions appear in the Model Studio console. Prices in other regions, including the Chinese mainland, are listed separately and are not reproduced here.
- Open checkpoints: hosted prices for the qwen3.6-27b and qwen3.8 open models were not in the portion of the pricing page read for this page; see the Qwen3.8 (Max and 27B) profile for Qwen3.8 hosted open-model prices.
What changed between models?
Only changes documented on the model cards and the pricing page are listed.
- Qwen3.6 (April 2026): hosted qwen3.6-plus and qwen3.6-flash with two input bands, plus two Apache 2.0 checkpoints. Qwen3.6-35B-A3B is a Mixture-of-Experts model with 256 experts, 8 routed and 1 shared active per token; Qwen3.6-27B is dense. Both are vision-language models with video input, thinking on by default, switchable off with enable_thinking, and a preserve_thinking option. The cards state the Qwen3-style /think and /nothink soft switches are not supported.
- Qwen3.6 to Qwen3.7 (May to July 2026): the generation includes a hosted Max tier (qwen3.7-max at $2.5 and $7.5; no qwen3.6-max appears in the pricing tables read), with a thinking-only preview snapshot first. qwen3.7-plus lowers list prices against qwen3.6-plus (for example $1.6 against $3 output in the 0 to 256K band) and qwen3.7-flash adds a 0 to 32K band at $0.030 input. No Qwen3.7 open checkpoint is named on the pages checked.
- Qwen3.7 to Qwen3.8 (August 2026): qwen3.8-max lists $2 and $6, lower than qwen3.7-max, and qwen3.8-flash uses a single 0 to 1M band. Open weights return, including a 2.4T total, 95B active checkpoint that the card calls the base of Qwen3.8-Max. The hosted Max adds vision input, non-thinking support, a 1M default context and official built-in tools over the open weights, which are text only and always think.
- Thinking controls in Qwen3.8: both Qwen3.8 cards add reasoning_effort with xhigh (default), medium and low, and turn preserve_thinking on by default. The 27B card notes that lower effort can increase total latency in multi-turn agent tasks.
- Output guidance: the Qwen3.6 cards suggest about 32,768 output tokens for most queries and up to 81,920 for complex maths and programming. The Qwen3.8 cards recommend up to 262,144 reasoning tokens and 131,072 final response tokens where the framework allows separate limits.
- Sampling defaults: Qwen3.6-35B-A3B lists presence_penalty 1.5 for general thinking tasks; Qwen3.8-2.4T-A95B lists presence_penalty 0.0, adjustable up to 2. Re-tune sampling settings when moving between generations.
For background on thinking modes and effort settings, see Reasoning Models Explained.
What workloads does Alibaba position each model for?
This section records the provider's own wording; it is not an assessment.
| Model | Positioning stated by provider | Source |
|---|---|---|
| Qwen3.8-Max | "A New Bar for Coding and Cowork" (citation title) | Qwen3.8-2.4T-A95B card |
| Qwen3.8-Flash | "Fast and cost-effective" | Model Studio models page |
| Qwen3.8-27B | Not stated separately; the card reuses the Qwen3.8-Max citation title | Qwen3.8-27B card |
| Qwen3.7-Plus | "Balanced performance, speed, and cost" | Model Studio models page |
| Qwen3.7-Max, Qwen3.7-Flash | Not stated by Alibaba on the pages checked (2026-11-15) | Not applicable |
| Qwen3.6-35B-A3B | "Agentic Coding Power, Now Open to All" (citation title) | Qwen3.6-35B-A3B card |
| Qwen3.6-27B | "Flagship-Level Coding in a 27B Dense Model" (citation title) | Qwen3.6-27B card |
| Qwen3.6-Plus, Qwen3.6-Flash | Not stated by Alibaba on the pages checked (2026-11-15) | Not applicable |
Worked cost example
Using the list prices above, 1,000 requests a day with a 100,000-token prompt and a 10,000-token response (reasoning included, since reasoning tokens bill as output) use 100 million input and 10 million output tokens. On qwen3.7-max that is $250 plus $75, $325 a day; on qwen3.8-max it is $200 plus $60, $260 a day. On qwen3.7-plus, every request stays in the 0 to 256K band: $40 plus $16, $56 a day; on qwen3.6-plus the same workload is $50 plus $30, $80 a day. Free quota and caching change the result, and token counts can differ between generations, so re-run the arithmetic on your own prompts. See Model Pricing Explained.
What is the context window of each model?
As of 2026-11-15. Sources: model cards for the open checkpoints; pricing tiers for the hosted models, whose maximum input band is the only context figure the pricing page gives.
| Model | Context window | Output guidance | Input | Output | Reasoning controls |
|---|---|---|---|---|---|
| Qwen3.8-Max (hosted) | 1M by default | Not stated by Alibaba on the pages checked (2026-11-15) | Text, image | Text | Thinking and non-thinking |
| Qwen3.8-Flash | 0 to 1M pricing band | Not stated by Alibaba on the pages checked (2026-11-15) | Not stated by Alibaba on the pages checked (2026-11-15) | Text | Not stated by Alibaba on the pages checked (2026-11-15) |
| Qwen3.8-2.4T-A95B | 262,144 native, up to 1,010,000 | Reasoning up to 262,144; final response up to 131,072 | Text | Text | Always thinking; reasoning_effort xhigh, medium, low |
| Qwen3.8-27B | 262,144 native, up to 1,000,000 | Reasoning up to 262,144; final response up to 131,072 | Text, image, video | Text | Thinking on by default, enable_thinking off; reasoning_effort xhigh, medium, low |
| Qwen3.7-Max | 0 to 1M pricing band | Not stated by Alibaba on the pages checked (2026-11-15) | Not stated by Alibaba on the pages checked (2026-11-15) | Text | Thinking and non-thinking (preview snapshot thinking only) |
| Qwen3.7-Plus | Up to 1M pricing band | Not stated by Alibaba on the pages checked (2026-11-15) | Text, image, video (listed under visual understanding) | Text | Thinking and non-thinking |
| Qwen3.7-Flash | Up to 1M pricing band | Not stated by Alibaba on the pages checked (2026-11-15) | Not stated by Alibaba on the pages checked (2026-11-15) | Text | Not stated by Alibaba on the pages checked (2026-11-15) |
| Qwen3.6-Plus, Qwen3.6-Flash | Up to 1M pricing band | Not stated by Alibaba on the pages checked (2026-11-15) | Not stated by Alibaba on the pages checked (2026-11-15) | Text | Not stated by Alibaba on the pages checked (2026-11-15) |
| Qwen3.6-35B-A3B | 262,144 native, up to 1,010,000 | About 32,768; up to 81,920 for complex tasks | Text, image, video | Text | Thinking on by default, enable_thinking off; preserve_thinking option |
| Qwen3.6-27B | 262,144 native, up to 1,010,000 | About 32,768; up to 81,920 for complex tasks | Text, image, video | Text | Thinking on by default, enable_thinking off; preserve_thinking option |
All four open checkpoints reach their extended windows through YaRN rope scaling, and the cards warn that static YaRN may reduce quality on shorter texts, so enable it only when long inputs need it. The hosted tiers apply long context for the user; on tiered models a prompt that crosses 256K moves the whole request into the higher price band. See Context Windows Explained.
Capabilities and tools compared
As of 2026-11-15. Architecture and serving facts from the model cards.
| Model | Architecture | Parameters | Serving frameworks stated by provider | Built-in tools |
|---|---|---|---|---|
| Qwen3.8-2.4T-A95B / Max | 92 layers; 3:1 Gated DeltaNet and Gated Attention, each followed by MoE; 512 experts, 10 routed plus 1 shared active | 2.4T total, 95B active | SGLang, vLLM, TokenSpeed; Transformers | Hosted Max: "official built-in tools" (list not stated) |
| Qwen3.8-27B | 64 layers, hidden 5120; 3:1 Gated DeltaNet and Gated Attention with FFN | 27B (Hugging Face shows 28B) | Not stated in the text read for this page | Not stated by Alibaba on the pages checked (2026-11-15) |
| Qwen3.7 hosted tiers | Not stated by Alibaba on the pages checked (2026-11-15) | Not stated by Alibaba on the pages checked (2026-11-15) | Model Studio only | Not stated by Alibaba on the pages checked (2026-11-15) |
| Qwen3.6-35B-A3B | 40 layers, hidden 2048; 3:1 Gated DeltaNet and Gated Attention, each followed by MoE; 256 experts, 8 routed plus 1 shared active | About 35B total, 3B active (Hugging Face shows 36B) | SGLang 0.5.10 or later, vLLM 0.19.0 or later, KTransformers, Transformers | Qwen-Agent and Qwen Code named for agentic use |
| Qwen3.6-27B | 64 layers, hidden 5120; 3:1 Gated DeltaNet and Gated Attention with FFN | 27B (Hugging Face shows 28B) | Not stated in the text read for this page | Not stated by Alibaba on the pages checked (2026-11-15) |
Every open checkpoint here uses the same 248,320-token padded vocabulary and multi-token prediction. The 27B dense models in both generations list identical layer counts, hidden size, head counts and FFN size, so changes between them come from training rather than shape. Fine-tuning services: not stated by Alibaba on the pages checked (2026-11-15). See Mixture of Experts Explained.
Where is each model available?
| Model | Availability stated by provider |
|---|---|
| Qwen3.8-Max | Model Studio (pricing page); Qwen Cloud named on the card |
| Qwen3.8-Flash | Model Studio |
| Qwen3.8 open checkpoints | Hugging Face; the 27B card lists a hosted version on Qwen Cloud as "coming soon" |
| Qwen3.7-Max, Plus, Flash | Model Studio |
| Qwen3.6-Plus, Flash | Model Studio |
| Qwen3.6 open checkpoints | Hugging Face; Model Studio (DashScope) per the cards; DeepInfra listed as an inference provider |
Prices on this page are for the Singapore deployment. Model availability differs by region, so confirm the model is offered in the region you deploy to. Self-hosting an open checkpoint moves patching and access control to the organisation; see the Self-Hosted LLM Security Guide.
Model cards and system cards
| Model | Model card or page | System card | Technical report |
|---|---|---|---|
| Qwen3.8-Max / 2.4T-A95B | Hugging Face card | Not stated by Alibaba on the pages checked (2026-11-15) | Not stated by Alibaba on the pages checked (2026-11-15) |
| Qwen3.8-27B | Hugging Face card | Not stated by Alibaba on the pages checked (2026-11-15) | Not stated by Alibaba on the pages checked (2026-11-15) |
| Qwen3.7 hosted tiers | Model Studio pricing and models pages | Not stated by Alibaba on the pages checked (2026-11-15) | Not stated by Alibaba on the pages checked (2026-11-15) |
| Qwen3.6-35B-A3B | Hugging Face card | Not stated by Alibaba on the pages checked (2026-11-15) | Not stated by Alibaba on the pages checked (2026-11-15) |
| Qwen3.6-27B | Hugging Face card | Not stated by Alibaba on the pages checked (2026-11-15) | Not stated by Alibaba on the pages checked (2026-11-15) |
See Model Cards and System Cards Explained.
Benchmarks
As of 2026-11-15. All figures are provider-reported on the Qwen Hugging Face cards. Only rows from the same table on the same card are paired, because suites and harnesses differ between cards. Independent leaderboard results (LMArena, Artificial Analysis, Epoch AI) were not checked for this page.
Qwen3.8-27B vs Qwen3.7-Plus vs Qwen3.6-27B (Qwen3.8-27B card)
| Benchmark | Qwen3.8-27B | Qwen3.7-Plus | Qwen3.6-27B |
|---|---|---|---|
| Terminal Bench 2.1 (Terminus) | 73.0 | 64.0 | 63.4 |
| SWE-bench Pro | 61.7 | 57.6 | 53.5 |
| GPQA Diamond | 89.2 | 90.3 | 87.8 |
| LiveCodeBench v6 | 90.3 | 89.6 | 83.9 |
| HLE | 30.8 | 34.7 | 24.0 |
| IFBench | 79.5 | 79.1 | 69.1 |
Qwen3.8-Max vs Qwen3.7-Max (Qwen3.8-2.4T-A95B card)
| Benchmark | Qwen3.8-Max | Qwen3.7-Max |
|---|---|---|
| Terminal Bench 2.1 | 86.6 | 74.5 |
| SWE-bench Pro | 67.7 | 60.6 |
| DeepSWE 1.1 | 56.6 | 21.6 |
| GPQA Diamond | 92.6 | 92.4 |
| HLE / HLE with tools | 43.6 / 56.2 | 41.4 / 53.5 |
| IFBench | 82.8 | 79.1 |
| MRCR v2 256K (8-needle) | 92.9 | 86.7 |
| LongBench v2 | 66.3 | 65.3 |
The Qwen3.8-2.4T-A95B card reports these in a Qwen3.8-Max column and notes that several results come from Qwen's own harnesses or in-house benchmarks. The Qwen3.6-35B-A3B card reports SWE-bench Verified 73.4 and Terminal-Bench 2.0 51.5, on suite versions the Qwen3.8 cards do not use, so those figures are not paired here. See How to Read LLM Benchmarks.
Deprecations and retirements
As of 2026-11-15. Source: Model Studio model deprecation page.
| Model | API ID | Notice or deprecation date | Retirement or shutdown date | Stated replacement |
|---|---|---|---|---|
| Qwen3.8 hosted models | qwen3.8-max, qwen3.8-max-0902, qwen3.8-flash | None named on the page | None named | None stated |
| Qwen3.7 hosted models | qwen3.7-max, -plus, -flash and dated snapshots | None named on the page | None named | None stated |
| Qwen3.7-Max preview | qwen3.7-max-preview, qwen3.7-max-2026-05-17 | None named; preview models get at least 14 days' notice | None named | None stated |
| Qwen3.6 hosted models | qwen3.6-plus, qwen3.6-flash and dated snapshots | None named on the page | None named | None stated |
The deprecation page lists batches by date (one scheduled for 19 January 2027, six deprecated on 10 October 2026) and puts the model IDs in linked announcements rather than on the page itself, so check those announcements and the Singapore region's Model Monitoring page for your own usage. Alibaba states that stable models receive a retirement notice no later than 3 months before end of service, and that accounts that called a retiring model in the last 3 months are notified. The Hugging Face checkpoints are not covered by this page's lifecycle notices. See the AI Model Retirement Guide and the LLM Deprecation and Retirement Tracker.
Data handling and security
The Model Studio security and privacy page (last updated 11 October 2026) states that Alibaba Cloud will never use customer data for model training, that Model Studio stores data generated from model and application calls in line with relevant laws, with terms set in the related agreements, that data transmitted during application building and model training is encrypted with AES-256, and that Model Studio has achieved SOC 2 compliance covering security, availability and confidentiality. The same terms apply across all three generations on the hosted service; the deployment region decides where calls are served.
Open checkpoints shift the data question to the organisation running them. The main operational risk across versions is the API credential: a Model Studio key can call every model the account can reach. See the LLMjacking Guide, the Open-Weight Model Security Guide and the Alibaba Cloud Security Guide.
How to choose a model: factual criteria
| Criterion | What the documentation shows |
|---|---|
| Open weights needed | Qwen3.6 (35B-A3B, 27B) and Qwen3.8 (2.4T-A95B, 27B); none named for Qwen3.7 |
| Licence | Apache 2.0 for Qwen3.6-35B-A3B, Qwen3.6-27B and Qwen3.8-27B; qwen3.8-max licence for Qwen3.8-2.4T-A95B |
| Max tier list price | qwen3.8-max $2 / $6; qwen3.7-max $2.5 / $7.5 |
| Long prompts on Plus and Flash | Qwen3.6 and Qwen3.7 Plus and Flash charge higher rates above 256K (or 32K for qwen3.7-flash); qwen3.8-flash uses one 0 to 1M band |
| Thinking-only behaviour | qwen3.7-max-preview and the Qwen3.8-2.4T-A95B weights cannot run without thinking |
| Effort control | reasoning_effort documented on the Qwen3.8 cards only |
| Self-hosting footprint | From 27B dense or 3B active (Qwen3.6-35B-A3B) to 2.4T total, 95B active |
| Snapshot pinning | Dated snapshots exist for every hosted Qwen3.6 and 3.7 model and for qwen3.8-max |
Run your own evaluation on representative prompts before switching generations. For a cross-provider view, see DeepSeek vs Qwen vs Kimi vs GLM and Open-Weight Frontier Models Compared.
Practitioner checklist
- Pin production calls to a dated snapshot ID (for example qwen3.7-plus-2026-05-26 or qwen3.8-max-0902) and record it in the AI inventory.
- Re-estimate cost when prompts cross a pricing band such as 256K tokens on Qwen3.6 and Qwen3.7 Plus and Flash.
- Re-tune sampling and output limits when moving from Qwen3.6 to Qwen3.8 checkpoints, including presence_penalty and the larger reasoning budget.
- Set reasoning_effort explicitly on Qwen3.8 rather than relying on the xhigh default where latency matters.
- Avoid building production flows on preview snapshots, which carry only 14 days' minimum retirement notice.
- Check the deprecation page announcements and Model Monitoring each quarter, including the batch scheduled for 19 January 2027.
- Read each checkpoint's licence file before redistribution or commercial use.
- Store Model Studio API keys in a secrets manager and set spend limits; see How to set spend limits on LLM API keys.
Standards and references
- 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 and privacy (accessed 2026-11-15)
- 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: Qwen3.6-35B-A3B model card (accessed 2026-11-15)
- Qwen Team: Qwen3.6-27B model card (accessed 2026-11-15)
Related NHI Mgmt Group resources: Alibaba Qwen Models, Qwen3.8 (Max and 27B), DeepSeek vs Qwen vs Kimi vs GLM, Frontier AI Models Compared, Chinese vs US Frontier Models Compared.