📋 Executive Summary
I began this Qwen AI model review three days after Alibaba unveiled Qwen3.8-Max, a 2.4-trillion-parameter system whose technical ambition is easier to verify than its commercial terms. The launch matters because Qwen is no longer a single chatbot or a neat open-model alternative. It is an expanding stack of hosted flagships, downloadable language models, multimodal systems, coding agents, speech tools, image models, embeddings and deployment frameworks. That breadth creates genuine value, but it also makes a simple verdict difficult.
The strongest answer is that Qwen is one of the most credible model families for organisations that want capability without accepting one vendor’s interface as the whole product. Its open-weight releases give developers unusual control over hosting, fine-tuning, latency and data location. Its Alibaba Cloud endpoints add an easier managed route, with OpenAI-compatible and, for selected services, Anthropic-compatible interfaces. The newest Max models aim higher, especially in agentic coding, long-context reasoning and multimodal work.
The weakness is fragmentation. Qwen3.8-Max was newly announced when this review closed on 6 August 2026, while Alibaba’s public pricing documentation still centred on Qwen3.7, Qwen3.6 and established specialist models. Model names, deployment scopes and promotional discounts vary by region. Open weights do not mean open training data, and a one-million-token context does not guarantee useful attention across a million tokens. This review therefore separates confirmed specifications from launch claims, compares the models by job rather than prestige, maps current costs, examines deployment bottlenecks and explains where Qwen is a better fit than Claude, Gemini, DeepSeek or a smaller local model.
What Qwen Has Become in 2026
Qwen began as Alibaba Cloud’s general-purpose large language model family, but the 2026 product shape is closer to an operating system for AI development. The core text line now sits beside vision-language models, omni-modal models, coding variants, image generation, text-to-speech, automatic speech recognition, embeddings, rerankers and agent tooling. Qwen Studio offers a consumer and developer interface. Model Studio exposes hosted APIs. Qwen Code brings an agent into the terminal, while Qwen Agent adds function calling, Model Context Protocol support, retrieval, code execution and application scaffolding.
That breadth is strategically important. A team can prototype with a hosted endpoint, move a stable workload to an open checkpoint, keep specialist media tasks on separate Qwen services and preserve a largely compatible application layer. The official Qwen3.6 repository lists Transformers, vLLM, SGLang, KTransformers and related ecosystems, while the broader Qwen documentation also supports llama.cpp, Ollama and LM Studio for smaller or quantised models. This is more deployment choice than most closed assistants provide.
The model family also reflects a larger market shift. Nathan Lambert, a researcher focused on open models, described the Qwen3 family as “the new default open models” in his 2025 year-end assessment for 2026. That judgement is not a benchmark score, but it captures adoption logic: developers choose a family when checkpoints, licences, runtimes, examples and community support align. Our overview of the early-2026 LLM landscape explains why open-weight systems increasingly compete through ecosystem reach rather than a single chat ranking. Qwen’s central promise is therefore optionality. Its central risk is that optionality transfers more engineering responsibility to the user.
Qwen Model Lineup and Technical Architecture
The Qwen name covers models with very different ownership, compute and deployment characteristics. The first purchasing mistake is to treat Max, Plus, open-weight MoE and specialist models as interchangeable. The second is to confuse total parameter count with the work performed for each token. Mixture-of-experts models store many expert networks but route each token through only a subset, which can reduce inference compute while preserving a large capacity envelope.
| Model or Family | Access | Confirmed Technical Position | Best-Fit Workload | Important Constraint |
| Qwen3.8-Max | New flagship; launch access varies | 2.4T total parameters, about 95B active, multimodal, up to 1M context reported at launch | Frontier reasoning, complex agents, very long multimodal inputs | Full public pricing and technical report not confirmed by 6 Aug 2026 |
| Qwen3.7-Max | Hosted proprietary API | Thinking and non-thinking modes; up to 1M input tier in pricing docs | High-end agents, coding and difficult reasoning | Regional prices and aliases differ |
| Qwen3.7-Plus | Hosted proprietary API | Tiered pricing to 1M input; lower cost than Max | Production assistants, retrieval, business automation | Requests above 256K enter a higher price tier |
| Qwen3.6-35B-A3B | Open weight, Apache 2.0 | 35B total, about 3B active; native 262,144 context, extensible to 1,010,000 | Agentic coding, multimodal local or private deployment | Checkpoint is about 71.9 GB before quantisation |
| Qwen3.6-27B | Open weight, Apache 2.0 | Dense 27B model in the Qwen3.6 open series | Coding, reasoning and predictable dense deployment | Higher active compute than the sparse 35B-A3B |
| Qwen3.5-397B-A17B | Open weight, Apache 2.0 | Native multimodality; 397B total and 17B active | Multilingual multimodal agents and research | Large storage and multi-GPU operational burden |
| Qwen Specialist Models | Hosted and/or open, model dependent | Vision, omni, coding, image, speech, embedding and reranking lines | Narrow production tasks | Capabilities, billing units and regional availability vary |
Qwen3.6-35B-A3B is the architectural sweet spot in this review. Its model card lists 256 experts, with eight routed experts plus one shared expert active in each MoE layer. The checkpoint’s 71.9 GB file size matters for download, storage and loading, even though only about 3 billion language parameters are activated per token. This is a practical information-gain point: sparse activation can lower per-token compute, but it does not eliminate memory, networking or model-loading costs. For teams comparing it with the proprietary systems covered in our Claude AI review, the real choice is not simply open versus closed. It is control versus operational simplicity.
Qwen3.8-Max Early Verdict: Impressive, but Not Yet Fully Auditable
Qwen3.8-Max is the obvious headline and the hardest model to review responsibly. Reuters reported on 3 August 2026 that Alibaba’s new flagship contains 2.4 trillion total parameters and activates about 95 billion per request through a mixture-of-experts design. The report also described multimodal input and a one-million-token context. Those numbers position Qwen3.8-Max as a scale-first response to the newest Chinese and US frontier systems.
The early technical signal is credible enough to take seriously. A model that activates roughly four per cent of its total parameters can potentially combine broad capacity with lower inference cost than a dense 2.4-trillion-parameter network. Public Qwen Code issue reports also identify a qwen3.8-max-preview endpoint with image and video modalities and a one-million-token context setting. Those reports are useful operational evidence, but they are not a substitute for an official model card, release paper, stable API documentation or independent evaluation.
The missing information changes the verdict. Alibaba’s Model Studio pricing page, updated 15 July, gives detailed rates for Qwen3.7-Max, Qwen3.7-Plus, Qwen3.6 open models and specialist services, but it did not provide a confirmed Qwen3.8-Max matrix when this review closed. Training data composition, safety evaluation, exact serving regions, rate limits and the final open-weight release terms also required clearer primary documentation. Qwen3.8-Max may become the family’s strongest option, but a production buyer should treat it as an evaluation candidate rather than a default procurement decision.
The caution is not anti-Qwen. It is the same discipline applied in the Gemini model review: a launch claim becomes decision-grade only when the model identifier, price, limits, safety controls, benchmark method and support terms are stable. The practical verdict today is promising capability, incomplete auditability.
Performance Evidence and the Benchmark-to-Production Gap
Alibaba’s official material supports a strong performance case, particularly for coding, reasoning and agents. At its May 2026 Qwen Conference, Alibaba Cloud said Qwen3.7-Max scored 56.6 on Artificial Analysis’s Intelligence Index, ranking fifth globally and first among Chinese models at that time. The Qwen3.6 release emphasised repository-level coding, frontend workflows and preserved reasoning context across conversations. Qwen3.5 claimed broad gains across language, code, agents, image and video understanding, including support for 201 languages and dialects.
| Evidence | What It Supports | Confidence | Why It Is Not the Final Verdict |
| Artificial Analysis score cited by Alibaba | Qwen3.7-Max was globally competitive in May 2026 | Medium-high | Third-party index, but cited through vendor press material and subject to model updates |
| Qwen3.6 official model card | Architecture, context, framework compatibility and release focus | High | Does not independently verify every benchmark claim |
| Qwen3.8-Max Reuters launch report | Scale, active parameters, modalities and context claim | Medium-high | Very recent release with limited independent production evidence |
| Vendor coding and reasoning benchmarks | Relative strength on selected tasks | Medium | Prompting, sampling, tool harness and test selection can change outcomes |
| Public GitHub issues | Real integration failures and edge cases | Medium for the reported setup | Community reports may be version-specific or unresolved |
| Internal production pilot | Latency, cost, reliability and task success for one organisation | Highest for that use case | Requires representative data, controls and repeatable scoring |
Benchmarks are most valuable as filters, not purchase orders. Coding scores can change when an agent receives a better repository index, tool schema or retry policy. Long-context tests may reward retrieval of a planted fact without measuring whether the model notices contradictions across a real contract set. Multilingual averages can hide weaker performance in low-resource domains. Even a stable model can perform differently across hosted providers because of quantisation, system prompts, token limits and safety layers.
A serious evaluation should therefore follow the principle in our AI accuracy investigation: measure the error that matters to the workflow, not the average score that looks best in a launch chart. For Qwen, that means task-completion rate, tool-call validity, citation accuracy, latency percentiles, tokens per successful outcome and human correction time.
Coding and Agentic Workflows
Coding is Qwen’s most persuasive practical category. Qwen3.6 was explicitly shaped by community feedback to improve frontend work and repository-level reasoning. Qwen Code provides a terminal agent, and the wider ecosystem supports OpenAI-compatible requests, function calling, MCP-style tools and common serving frameworks. Alibaba’s May conference positioned Qwen3.7-Max as a foundation for agents rather than only a chat model.
Dr Feifei Li, Alibaba Cloud’s CTO and President of International Business, called the agentic era “a paradigm shift in how we interact with technology”. The phrase is promotional, but the engineering implication is correct: an agent model must do more than produce plausible text. It must select a tool, construct valid arguments, observe results, preserve state, recover from errors and stop before it causes damage. Qwen’s model family, Qwen Agent and Qwen Code give developers many of those building blocks.
The weak point is harness sensitivity. Public Qwen Code issues reported that a Qwen3.8 preview endpoint rejected internal operations when the client sent enable_thinking=false to a thinking-only model. Another issue described relative file paths resolving to a temporary directory rather than the working directory. These are not reasons to reject Qwen, but they show why model intelligence and agent reliability are different products. A correct answer can still produce the wrong file path, repeat a cancelled tool or fail during context compression.
Teams considering Qwen for autonomous work should compare this with the architecture in the DeepSeek agent guide: use typed tools, explicit permissions, idempotent writes, a hard step budget, durable external state, structured validation and human approval before irreversible actions. Qwen is strong enough to justify that investment. It is not strong enough to make the controls optional.
Multimodal Input and Long-Context Reasoning
Qwen’s multimodal direction is broader than image question answering. Qwen3.5 was released as a natively multimodal model trained across text, images, video, STEM and reasoning data. Alibaba says it can analyse videos up to two hours, convert hand-drawn interface sketches into frontend code, interpret screens and support visual agents. Qwen3.6-35B-A3B also includes a vision encoder, and the Qwen3.8-Max launch was reported to support text, images and video.
The long-context numbers are equally ambitious. Qwen3.6-35B-A3B lists a native 262,144-token context that can be extended to 1,010,000. Model Studio prices Qwen3.7 variants in tiers reaching one million input tokens. Yet context capacity is not context quality. A model can technically accept a huge prompt while losing chronology, overweighting recent instructions, missing a small exception or becoming too expensive for repeated use.
For document work, the best design is usually retrieval plus selective long context. Parse source files, preserve page and section metadata, retrieve a narrow evidence set, and reserve the largest window for cases where relationships across the whole collection genuinely matter. For video, sample at a task-appropriate rate and retain timestamps so a reviewer can return to the evidence. For interface agents, separate visual interpretation from the authority to click, buy, delete or send.
This matters because multimodal confidence can disguise ordinary fabrication. The AI hallucinations explainer describes how fluent specificity can remain unsupported. Qwen’s large context and visual capabilities improve the amount of evidence a system can consider, but they do not remove the need for provenance, contradiction checks and human review.
API Pricing, Free Quotas and Hidden Cost Multipliers
Alibaba Cloud’s pricing is competitive, but it is not one simple global number. Deployment scope matters. International, global and Chinese-mainland services can list different rates, discounts and model aliases. Some offers are temporary. Requests that cross a context threshold can move every token in the request into a more expensive tier. Batch inference and context caching can lower cost, but the two discounts may not combine.
| Model | Scope and Input Band | Input per 1M Tokens | Output per 1M Tokens | Confirmed Cap or Note |
| Qwen3.7-Max | Global, up to 1M | $1.65 list | $4.951 list | Thinking and non-thinking; regional promotional discounts may apply |
| Qwen3.7-Plus | International, 0-256K | $0.40 list | $1.60 list | 1M-token free quota for 90 days after activation |
| Qwen3.7-Plus | International, 256K-1M | $1.20 list | $4.80 list | All request tokens billed at the higher tier |
| Qwen3.6-35B-A3B | International, 0-256K | $0.375 | $2.25 | Open-weight model also available for self-hosting |
| Qwen3.6-27B | International, 0-256K | $0.60 | $3.60 | Dense model; 1M-token free quota for 90 days |
| Qwen Max legacy alias | International, standard | $1.60 | $6.40 | Non-thinking mode; 50% batch discount listed |
| Qwen3.8-Max | Not publicly confirmed at review close | Not confirmed | Not confirmed | Do not infer price from Qwen3.7 or preview plans |
Caching deserves special attention. Alibaba’s pricing documentation says an explicit cache can cost 125 per cent of standard input pricing to create, while a cache hit may cost 10 per cent of standard input pricing. That can be attractive for a repeated system prompt, policy corpus or stable codebase, but only if the cache-hit rate is high enough. Cache invalidation, regional availability and privacy policy still matter.
The hosted comparison is also incomplete without self-hosting cost. The Perplexity and DeepSeek comparison shows why a cheap token price can still lose to a managed product once search, storage, orchestration, observability and review are included. Qwen’s best cost advantage appears when a team can route simple work to Plus or an open model, reserve Max for difficult cases and measure cost per accepted result.
Local Deployment, Hardware and Throughput Reality
Open weights are Qwen’s clearest differentiator, but downloading a checkpoint is not the same as operating a service. Qwen3.6-35B-A3B occupies about 71.9 GB in its published Hugging Face repository before local caches and runtime overhead. Full-precision or high-quality low-bit deployment therefore needs serious memory planning. The 397B-A17B Qwen3.5 flagship is a multi-node or high-end multi-GPU proposition for most organisations.
| Deployment Pattern | Suitable Qwen Model | Why It Works | Primary Bottleneck |
| Developer workstation | Smaller quantised Qwen3 or specialist model | Private experimentation and low request volume | RAM or VRAM, generation speed and quantisation quality |
| Single high-memory GPU server | Quantised 27B or 35B-A3B | Good balance for internal coding, retrieval and multimodal tasks | KV cache at long context and concurrent users |
| Multi-GPU inference cluster | 35B-A3B, 397B-A17B or larger variants | Higher throughput, larger contexts and production concurrency | Tensor parallelism, networking, model loading and utilisation |
| Managed Model Studio API | Qwen3.7 Max or Plus, specialist APIs | Fast launch, elastic capacity and simpler operations | Regional price, data policy, rate limits and vendor dependence |
| Hybrid router | Local open model plus hosted Max | Keeps routine or sensitive work local and escalates hard tasks | Routing quality, duplicate infrastructure and audit complexity |
Alibaba’s 2026 RTP-LLM paper is relevant because it addresses the parts often hidden behind a model benchmark: loading, prefill-decode separation, KV-cache reuse, speculative decoding, quantisation and multimodal scheduling. The authors reported deployment across Alibaba serving more than 100 million users and measured substantial improvements over baseline inference engines. Those results are not a guarantee for a reader’s cluster, but they show that production efficiency comes from the serving system as much as the checkpoint.
The open-weight debate also extends beyond cost. NVIDIA chief executive Jensen Huang argued in July 2026 that “we need open weight for security”. That view fits Qwen’s local option, but security is not automatic. Teams should follow the controls in the AI privacy concerns guide: verify model provenance, isolate runtimes, patch serving software, restrict outbound network access and log every tool action.
A Production Implementation Workflow
A dependable Qwen implementation starts with workload evidence, not a model name. The following workflow is deliberately conservative because it makes capability, cost and risk visible before scale.
1. Define one measurable job. Specify the input, permitted tools, acceptable output, failure modes and human owner. Example: classify support tickets and draft a response, but never send it.
2. Build a fixed evaluation set. Use representative English, Chinese and domain-specific examples, including ambiguous requests, malicious instructions, long documents and tool errors.
3. Test three model tiers. Compare Qwen3.7-Plus, Qwen3.7-Max and an open Qwen3.6 checkpoint. Keep prompts, temperature, tool schemas and scoring constant.
4. Measure outcome economics. Record input and output tokens, latency, retries, cache hits, human correction time and infrastructure cost per accepted result.
5. Add retrieval and provenance. Store source identifiers beside every retrieved chunk. Require the model to return evidence IDs for factual claims.
6. Constrain tools. Use server-side allow-lists, typed arguments, idempotency keys, timeouts, step limits and approval gates. Never rely on a system prompt as the permission boundary.
7. Canary the release. Route a small traffic share, compare against the current system, monitor drift by model identifier and keep a rollback path.
The hosted API can be approached through the OpenAI client pattern shown in Qwen’s own examples. The base URL and exact model ID must match the selected deployment region.
from openai import OpenAI
import os
client = OpenAI(
api_key=os.environ[“DASHSCOPE_API_KEY”],
base_url=”https://dashscope-intl.aliyuncs.com/compatible-mode/v1″,
)
response = client.chat.completions.create(
model=”qwen3.7-plus”,
messages=[{“role”: “user”, “content”: “Return valid JSON for this ticket.”}],
response_format={“type”: “json_object”},
)
print(response.choices[0].message.content)
Production code should add timeouts, retry classes, schema validation, token budgets and model-version logging. Anthropic-compatible endpoints and thinking controls require version-specific testing. Do not assume every Qwen alias supports the same reasoning switch, tool-choice behaviour or side-query path.
Known Limitations, Governance and Safety
Qwen’s limitations fall into four categories: disclosure, behaviour, operations and governance. Disclosure is uneven across the family. Open checkpoints provide weights and model cards, but not necessarily the complete training dataset, data-removal process or every safety evaluation. Proprietary Max releases may appear before detailed pricing or technical reports. Buyers should distinguish open weight from fully open science.
Behavioural limits include hallucination, prompt injection, tool-argument mistakes and multilingual inconsistency. A public Qwen3.6 issue reported a simple prompt-injection failure, while other issues described language-specific quality and multi-turn system-message problems. Issue trackers are not controlled benchmarks, yet they are valuable because they reveal the shape of failures that polished demos omit. Every organisation should reproduce the relevant failure class in its own harness.
Operationally, long context increases KV-cache memory and latency. MoE reduces active compute but can raise routing, communication and loading complexity. Model aliases can point to dated snapshots, regional services can differ, and temporary discounts can distort a cost model. Preserved reasoning context may reduce repeated work, but any retained or replayed hidden state requires careful treatment of sensitive information and auditability.
Governance depends on where the service runs, what the contract says and whether the model can take action. A self-hosted Qwen model can improve data locality, but the operator becomes responsible for access control, monitoring, patching, abuse prevention and incident response. A hosted endpoint simplifies operations, but requires a review of data processing, retention, residency, subcontractors and model-improvement terms. Neither route removes the need for human accountability.
The balanced lesson from our DeepSeek V4 launch analysis applies here: low cost and open access can accelerate adoption while also increasing the number of poorly governed deployments. Qwen should be evaluated as infrastructure, not only as an assistant.
Qwen vs Claude, Gemini and DeepSeek: Who Should Choose It?
Qwen is strongest when deployment freedom is a first-order requirement. A multilingual product team that needs Chinese and English performance, private hosting, quantisation, fine-tuning and a compatible managed fallback has a compelling reason to shortlist it. A coding team can pair Qwen Code or Qwen Agent with Qwen3.6 locally and escalate difficult tasks to Qwen3.7-Max. A media workflow can stay within the wider Qwen family for vision, speech, image and embedding services.
Claude remains attractive when a team wants a polished closed-model experience, strong long-document work and mature coding behaviour without operating weights. Gemini has advantages where Google Workspace, search, Android or cloud integration already defines the environment. DeepSeek competes aggressively on token cost and reasoning, but its product, privacy and deployment trade-offs differ. Qwen’s advantage is not that it wins every task. It is that the family offers more routes between local control and hosted scale.
Choose Qwen when the organisation has engineers who can benchmark models, secure a runtime, manage prompts and tools, and monitor version changes. Choose a managed competitor when time-to-value, vendor support, standardised governance and a single integrated interface matter more than portability. Choose a smaller local model when the task is narrow enough that frontier capability would merely increase cost and risk.
Eddie Wu, Alibaba Group’s chief executive, said in May 2026 that the company’s AI investments had moved “from incubation to commercialization at scale”. The business evidence supports that claim: Alibaba reported 40 per cent growth in external cloud revenue and an eightfold increase in the Model Studio customer base year over year. Commercial momentum makes Qwen durable, but it does not settle the model choice for an individual buyer.
The final recommendation is a structured pilot. Test Qwen3.7-Plus as the cost baseline, Qwen3.7-Max or Qwen3.8-Max as the difficult-task candidate once terms are confirmed, and Qwen3.6-35B-A3B as the control-oriented open model. Keep the winner only if it improves accepted outcomes after infrastructure, review and risk costs are included.
Our Research Methodology
This review was completed on 6 August 2026 using a source hierarchy designed for a fast-moving model launch. Specifications and prices were checked first against Alibaba Cloud Model Studio documentation, official Qwen GitHub repositories, Qwen model cards on Hugging Face, Qwen release posts and Alibaba Group or Alibaba Cloud announcements. The Qwen3.8-Max launch details were cross-checked against Reuters because a complete official pricing matrix and technical report were not publicly confirmed when the review closed.
The comparison focused on five decision metrics: confirmed capability, deployment control, context and multimodality, total cost, and operational reliability. Benchmark claims were classified by evidence type rather than copied into one league table. Official architecture details received higher confidence than vendor-selected performance claims. Public GitHub issues were used only as examples of reproducible integration risks, not as evidence that every deployment fails in the same way.
No paid Qwen3.8-Max API account or multi-GPU local benchmark was available for a controlled, repeatable inference test. The article therefore does not claim first-hand latency, quality or cost measurements for that new model. Where evidence was incomplete, the text states the limitation. Pricing was captured from the public documentation dated 15 July 2026 and should be rechecked before procurement because regional promotions, aliases and deployment scopes can change.
Internal links were selected from live indexed Perplexity AI Magazine pages after the site’s XML sitemap endpoints did not return parseable content through the research layer. Eight semantically relevant links were used once each and distributed across body sections covering model reviews, open-weight competition, accuracy, hallucinations, privacy and agent engineering.
This article was researched and drafted with AI assistance and reviewed by the Sami Ullah Khan editorial desk at Perplexity AI Magazine. All data, citations, pricing figures, and named quotes have been independently verified against primary sources before publication.
After WordPress publication, the editorial team should run the required back-button and hidden-content checks. Navigation must return to the previous page without a redirect loop, and the rendered page should contain no search-visible text hidden through display, visibility, matching colour, zero font size or large negative positioning.
Conclusion
Qwen’s 2026 story is not simply that Alibaba built another powerful model. It is that one family now spans open weights, hosted flagships, coding agents, multimodal systems and production infrastructure. Qwen3.6-35B-A3B offers a particularly convincing balance of sparse compute, long context, Apache 2.0 licensing and broad runtime support. Qwen3.7 provides clearer hosted pricing and agent-focused capability. Qwen3.8-Max raises the ceiling again, but its launch-day scale is better documented than its complete commercial and technical terms.
That balance leads to a qualified positive verdict. Qwen belongs on the shortlist for multilingual products, coding systems, private deployments and teams that want to move between managed APIs and self-hosted checkpoints. It is less suitable for organisations that expect a model to arrive as a fully governed application with stable global pricing, uniform regional behaviour and minimal engineering work.
The open questions are important: whether Qwen3.8-Max will receive stable open weights, how independent evaluations will compare with launch benchmarks, what its final price will be, and how reliably its agent behaviour will hold across tool harnesses. Until those answers settle, the smartest Qwen strategy is not to chase the largest model. It is to route each workload to the smallest, best-documented Qwen model that meets a measured production threshold.
Frequently Asked Questions
Is Qwen AI Free to Use?
Qwen Studio offers free access subject to product limits, while many open Qwen checkpoints can be downloaded under Apache 2.0. Hosted Model Studio APIs are metered. Selected international models include introductory free token quotas for 90 days after activation, but eligibility, regions and promotions vary.
Is Qwen Open Source?
Many Qwen models are open weight under Apache 2.0, including Qwen3.6 checkpoints. That does not mean every flagship is open or that training data and the entire development process are public. Qwen3.7-Max is a hosted proprietary model, and Qwen3.8-Max terms required further confirmation at review close.
What Is the Best Qwen Model for Coding?
Qwen3.6-35B-A3B is a strong open-weight coding and agentic option, while Qwen3.7-Max targets more demanding hosted workflows. The best choice depends on repository size, tool harness, latency, privacy and cost. Test both on a fixed set of real issues.
Can Qwen Run Locally?
Yes. Open Qwen models can run through Transformers, vLLM, SGLang, llama.cpp, Ollama, LM Studio and other compatible runtimes. Hardware needs vary sharply by model size and quantisation. Large checkpoints require substantial VRAM, RAM, storage and operational expertise.
Does Qwen Support Images and Video?
Yes. Qwen3.5 is natively multimodal, Qwen3.6-35B-A3B includes a vision encoder, and Qwen3.8-Max was reported to accept text, images and video. Specialist Qwen-VL and Qwen-Omni models provide additional visual, audio and video capabilities.
How Large Is Qwen’s Context Window?
It depends on the model. Qwen3.6-35B-A3B lists 262,144 tokens natively and extension to about 1.01 million. Qwen3.7 pricing supports requests up to one million tokens. A large advertised window should still be tested for retrieval quality, latency and cost.
Is Qwen Better Than DeepSeek?
Qwen offers a broader family, stronger deployment variety and extensive multimodal and agent tooling. DeepSeek often competes aggressively on API cost and reasoning. Neither is universally better. Evaluate the exact model, region, privacy terms, tools and successful-outcome cost.
Is Qwen Safe for Business Data?
It can be used safely only with appropriate controls. Self-hosting improves locality but transfers security duties to the operator. Hosted use requires contract, residency, retention and access reviews. In both cases, restrict tools, minimise sensitive prompts, log actions and require human approval for consequential decisions.
APA References
1. Alibaba Cloud. (2026, July 15). Model inference pricing.
5. Qwen Team. (2026). Qwen3.6 official repository.
6. Qwen Team. (2026). Qwen3.6-35B-A3B model card.
7. Qwen Team. (2025, September 24). Qwen3-Max: Just scale it.
8. Tan, B., et al. (2026). RTP-LLM: High-performance Alibaba LLM inference engine. arXiv.
