Qwen 3.8 Max: The Definitive Evidence-Led Guide to Alibaba’s New Flagship Model

Alibaba’s production Qwen3.8-Max is a 2.4-trillion-parameter multimodal flagship with a one-million-token context window, aggressive $2/$6 API pricing and serious coding-agent ambitions. The launch case is strong. The independent case is still being built.

Evidence refreshed September 5, 2026. This is an evidence-led launch analysis, not a hands-on Kingy benchmark. We did not have an authenticated production API credential, so we did not run qwen3.8-max. Preview anecdotes and results are not presented as production testing.

The short verdict

Qwen3.8-Max is the most consequential Qwen launch in this generation because Alibaba has combined four things buyers rarely get together: frontier-scale sparse capacity, native visual and video input, a million-token context window and a hosted price materially below leading proprietary coding models. Its production API ID is qwen3.8-max; that is not the same model identity as July’s qwen3.8-max-preview.

The evidence supports calling Qwen3.8-Max a credible frontier-class hosted model. It does not yet support calling it the best model overall, a proven replacement for the strongest coding agents or a model suited to an ordinary local workstation. Alibaba’s benchmark pack is broad but methodologically mixed. Launch-day Arena data gives Qwen an encouraging independent frontend-code signal, while a broad production evaluation from an established independent lab was still missing at the cutoff. Alibaba has since released the official Qwen3.8-2.4T-A95B weights. The open model uses a custom qwen3.8-max license and differs from the hosted product: it is text-only, requires thinking and has a 262,144-token native context that can be extended to 1,010,000.

Kingy verdict

• What it is: Alibaba’s new production Max-tier mixture-of-experts model, officially styled Qwen3.8-Max, with 2.4T total parameters and 95B active per inference.
• What materially changed: A Qwen flagship now combines long-context coding and knowledge work with image and video input, production reasoning controls, built-in tools and five-region hosted access.
• Who should consider it now: Teams routing documents, screenshots, video and code through one hosted model; developers willing to benchmark a new API against their own agent workloads; buyers who find Kimi K3’s output price too high but need more than DeepSeek’s text-only path.
• Who should wait: Local-deployment teams without multi-node data-centre infrastructure, regulated buyers needing settled retention language, and anyone whose decision depends on independently reproduced repository-scale performance.
• Verdict confidence: Moderate. Identity, specifications, access and price are well documented; real-world quality and operational reliability are not yet independently established.
• Testing status: No direct Kingy run. All hands-on claims are excluded.

Qwen3.8-Max is not Qwen3.8-Max-Preview

The naming is easy to mishandle, especially while launch documentation is still propagating.

On July 19, Alibaba introduced qwen3.8-max-preview through its Token Plan. Qwen Code added that preview to its plan model list four days later. On August 3, Qwen announced the production model, Qwen3.8-Max, with the API ID qwen3.8-max. The official launch post calls this an official release; the Alibaba Cloud model catalog now lists the production ID across multiple regions and protocols.

That chronology matters. A preview result can reveal direction, but it cannot establish production quality, latency, tool reliability, pricing or privacy behavior. It also matters for integrations: a client pinned to qwen3.8-max-preview has not silently become version-pinned production evidence.

The model should not be confused with Qwen3-8B, Qwen3-Max, Qwen3-Max-Thinking, Qwen3.6-Max or the prior Qwen3.7 Max generation. Similar names do not imply shared weights, identical endpoints or transferable benchmarks.

Confirmed, claimed and still unknown

The cleanest way to read this launch is to separate product facts from evaluation claims and gaps.

What 2.4 trillion parameters actually means

Qwen3.8-Max is a mixture-of-experts model. Alibaba reports 2.4 trillion parameters in the full expert pool and 95 billion active during an inference step. The active share is about four per cent of the total.

That sparsity is the core economic idea: the model can hold far more learned capacity than it exercises for every token. It does not make Qwen3.8-Max a conventional 95B model. A serving system still needs the expert pool available, plus the attention layers, router, multimodal components, runtime workspace and KV cache. “Active parameters” describe a compute path, not a download size.

The raw arithmetic is sobering. The official BF16 checkpoint’s 2.4T parameters imply roughly 4.8 TB before overhead. Eight-bit storage would be about 2.4 TB; four-bit storage about 1.2 TB. The 95B active path alone represents roughly 190 GB at BF16. These are lower bounds, not Alibaba hardware requirements; Qwen’s official repository publishes BF16 weights and does not make an official quantized release claim.

This is why the open-weight promise should be interpreted precisely. The shipped Qwen3.8-2.4T-A95B files still define a data-centre-scale deployment. Local availability requires more than a download button: a license, serving recipe, supported engines, precision details, enough aggregate accelerator memory, a fast interconnect and measured quality under quantization. For ordinary developers, a smaller Qwen3.8 model could ultimately matter more than Max.

The official model card now discloses 92 layers, 512 experts, 10 routed experts plus one shared expert per token, an 8,192 hidden size and grouped-query attention. Training-token total remains undisclosed.

Context, multimodality and reasoning controls

The QwenCloud production page lists a one-million-token window, roughly 991.8K maximum input, 983.61K maximum thinking-mode input and 131.07K maximum output. Some launch integration examples set maxTokens to 65,536; that is a client configuration, not the documented platform ceiling.

The model accepts text, images and video and produces text. That gives it a plausible advantage for work that arrives as screenshots, scanned documents, diagrams, slide decks or video rather than clean text. But protocol details matter. Qwen’s Responses interface does not accept video input; official guidance routes video through Chat Completions or DashScope. A buyer should test the exact endpoint and payload path intended for production instead of treating “native video” as universal interface parity.

Thinking is the default. The Chat Completions reference exposes reasoning_effort values of low, medium and xhigh, with xhigh as the default. none disables thinking; max and high map to xhigh. Developers can instead set a thinking budget, but cannot combine the two controls.

Preserved thinking is enabled by default. In multi-turn and tool sessions, the historical reasoning_content must be passed back and is billed as input. That improves continuity, but it can quietly enlarge an agent loop’s context and cost. Cost tests should therefore report both visible output and reasoning-history tokens.

Qwen documents function calling, parallel tool calls, structured output, prefix completion, batches and caching. Its Responses API adds code interpreter, web extraction, web search and text-to-image or image-to-image search. Built-in tools have separate economics: QwenCloud currently lists web search at $10 per 1,000 calls and either image-search mode at $8 per 1,000. Code interpreter and web extraction were temporarily free at the cutoff, but still add model tokens.

The following is a documentation-derived request shape, not a Kingy-tested example. Use the current regional workspace endpoint rather than copying an old launch URL, and keep the returned assistant reasoning fields in later tool turns:

import os
from openai import OpenAI

client = OpenAI(
    api_key=os.environ["QWEN_API_KEY"],
    base_url=os.environ["QWEN_BASE_URL"],  # Current regional /compatible-mode/v1 URL
)

result = client.chat.completions.create(
    model="qwen3.8-max",
    messages=[{"role": "user", "content": "Audit this specification and return structured risks."}],
    max_completion_tokens=8192,
    extra_body={
        "reasoning_effort": "medium",
        "preserve_thinking": True,
    },
)

print(result.choices[0].message.content)

For production, replace the prose prompt with a versioned prompt fixture and a strict output schema; log region, model ID, reasoning setting, latency, input, cached-input, reasoning-output and visible-output tokens. A successful HTTP response is not an accepted task. Validate the structure and factual claims before committing the result to a workflow.

Access, interfaces and regional reality

Alibaba’s catalog lists Qwen3.8-Max in Beijing, Singapore, Tokyo, Frankfurt and US Virginia. Each entry exposes OpenAI-compatible, Anthropic-compatible and DashScope routes. The production launch also names QwenCloud and integrations including Qwen Code, Codex, Claude Code, Qoder CLI and OpenClaw.

Five region labels do not automatically mean five identical data-residency guarantees. Alibaba’s region and deployment-scope guide distinguishes the region receiving and storing a request from the scope in which inference can run. A Global scope may schedule across boundaries. Organizations that require US-, EU-, Japan-, international-ex-China or China-mainland processing must select the corresponding scope and confirm that the model, endpoint and contract match that requirement.

Privacy language is similarly plan-specific. General QwenCloud documentation says API prompts and completions are not used for training and ordinary inference is processed in memory. Responses stores state by default. QwenCloud’s zero-retention page says conversation data linked through previous_response_id is deleted after seven days; its Safety guide says store=true conversation data is retained for 30 days. The documentation does not explain whether these are different storage paths or which duration governs a default stored response. China Personal Token Plan authorizes content use for service improvement and model optimization; Team Token Plan says conversations are not used for training. The defensible enterprise position is not “Qwen never trains on customer data.” It is: identify the exact portal, plan, region, API and storage setting, then obtain contractual confirmation.

What the benchmark pack proves—and what it does not

Qwen’s launch table is unusually broad. On coding and agents, Alibaba reports 86.6 on Terminal-Bench 2.1, 67.7 on SWE-bench Pro, 56.6 on DeepSWE 1.1, 93.0 on PaperBench, 74.8 on CoWorkBench and 81.9 on WideSearch. It reports 92.6 on GPQA Diamond, 82.8 on IFBench and 92.9 on MRCR v2 at 256K. Multimodal results include 82.3 on MMMU-Pro, 86.1 on OSWorld Verified, 92.1 on OmniDocBench 1.5 and 90.4 on Video-MME with subtitles.

Those results support a narrow conclusion: Qwen3.8-Max is competitive across coding, research, tool use, long context, documents, computer use and video. They do not support a universal crown.

Qwen-reported production results—not independent Kingy tests

The adjacent methodology warning is the point: these are useful vendor disclosures, not a controlled neutral leaderboard. Scores should remain tied to Qwen’s launch configuration and date.

The table is not one controlled tournament. Qwen explicitly mixes Claude Code, Codex, OpenCode, Qwen-Agent and QwenWork harnesses. Some rival numbers are the best published results; others are Alibaba reruns. Timeouts, token ceilings, sampling counts, tool access and judges differ. Several benchmarks are Qwen-created or internally operated. Some multimodal tests allow code, search, subtitles, memory or other scaffolding. Alibaba also records manual corrections in parts of its evaluation process.

The model does not lead every row. Strong proprietary systems remain ahead on several demanding repository, agent and video tasks. For example, Qwen’s 67.7 SWE-bench Pro and 56.6 DeepSWE figures do not erase the higher scores reported for some frontier coding competitors under their own harnesses.

The independent picture is thinner. Arena’s preliminary WebDev snapshot scored Qwen3.8-Max at 1,668 ±18 and Kimi K3 at 1,676 ±12; the intervals overlap, making this a statistical tie. That is useful human-preference evidence for frontend code, not a general intelligence score or proof of repository-scale reliability. No completed production Qwen3.8-Max evaluation from Artificial Analysis was verified by the cutoff. This absence lowers confidence; it does not prove weakness.

Qwen3.8-Max versus the models buyers will actually consider

The nearest open-weight comparison is Kimi K3: another enormous multimodal MoE built for long-horizon agents. DeepSeek V4 Flash 0731 is the value and deployability contrast: much smaller, text-only, MIT licensed and radically cheaper. Qwen’s prior Max generation is the migration baseline, but launch-day buyers should test production qwen3.8-max, not infer its behavior from a Qwen3.7-labelled result.

Kimi K3’s official model card discloses 2.8T total and 104B active parameters, a 1M window, detailed routing and downloadable MXFP4 weights. It is stronger than Qwen on evidence maturity: Artificial Analysis scored its max configuration at 57. The trade-offs are a $3 input/$15 output API and a custom license with conditions for some large model-as-a-service and consumer deployments.

DeepSeek’s July 31 release needs exact naming. DeepSeek-V4-Flash-0731 is a re-post-trained model served behind the unchanged deepseek-v4-flash alias. April Preview scores should not be attached to 0731. The official 0731 model card documents 284B total and 13B active parameters, text-only input, a 1M window, 384K maximum output and MIT weights. Its API price is $0.14 input and $0.28 output per million tokens. Artificial Analysis scored the 0731 max configuration at 50. DeepSeek is the obvious value winner, but it is not a visual/video substitute for Qwen.

Pricing: cheap for its class, not simple

QwenCloud lists $2 per million input tokens and $6 per million output tokens, including reasoning. It separately lists $0.25 for implicit-cache input, $2.50 to create an explicit cache and $0.17 for explicit-cache reads. Regional Model Studio prices differ: Beijing, Virginia, Frankfurt and Tokyo list CNY 12 input/CNY 36 output, while Singapore lists CNY 14.988/CNY 44.965. Portal, region, currency and account type must accompany any price claim.

The cache documentation itself needs a warning label. Generic Alibaba guidance says explicit creation costs 125% of normal input, explicit reads 10% and implicit reads 20%. The model-specific QwenCloud dollar table implies different percentages. For budgeting, use the model-specific numbers shown in the account you will bill, then verify with a small controlled invoice test.

Here is a normalized request with 100K uncached input and 20K output:

That makes Qwen about half Kimi’s cost for this shape and about 16 times DeepSeek’s. The ranking can change at the task level. A verbose reasoner, a failed tool loop or a weak cache hit rate can overwhelm nominal token savings. Built-in Qwen web search and image search add per-call charges. Preserved reasoning history can expand later requests. The meaningful metric is cost per accepted result under the same prompts, tools, retry policy and review standard.

A cache-sensitive agent loop shows why the billing details matter. Assume ten calls, each with a 100K-token stable prefix, 5K new input and 10K output. Across the run, that is 900K potential cache-read tokens, 150K uncached input tokens and 100K output tokens. Using QwenCloud’s model-specific implicit-cache rate, the derived bill is about $1.125 if all nine repeated prefixes hit: $0.225 cached input, $0.30 uncached input and $0.60 output. With no cache hits, the same 1.05M input and 100K output cost about $2.70. The apparent saving is roughly 58%, but it depends on exact prefix stability and actual hit reporting.

Under the same idealized token shape, Kimi K3 costs about $2.22 using its $0.30 cache-read rate, while DeepSeek V4 Flash 0731 costs about $0.0515 using its $0.0028 cache-read rate. These are arithmetic comparisons, not measured task costs. Reasoning verbosity, cache invalidation, retries and differing success rates can dominate the bill.

Token Plan subscriptions and promotional credits belong in a separate budget line. They can be useful access mechanisms, but they are not durable pay-as-you-go unit prices and may carry different data terms.

Which model should you choose?

This is a routing shortlist, not a winner’s podium. Keep the harness, prompt, tools, reasoning budget and acceptance test fixed when moving between models.

Coding agents: Put Qwen on the shortlist, especially when code work includes screenshots, diagrams, tickets or documentation. Do not replace a proven coding model from a vendor benchmark table alone. Run a frozen repository suite with the same harness, tools, reasoning effort, timeout and retry budget. Track accepted patches, regressions, human repair time, latency and total billed tokens.

Research and document workflows: This is Qwen’s most compelling launch use case. The combination of long context, document/visual input, structured output, web extraction and search can reduce model routing. Test citation faithfulness, table extraction, contradictory-source handling and performance at increasing context lengths; a one-million-token maximum does not guarantee one-million-token accuracy.

Enterprise automation: The interface compatibility and regional catalog are attractive, but procurement has work to do. Confirm deployment scope, endpoint, retention, store behavior, training-use terms, support, rate limits and incident process in writing. QwenCloud lists 15,000 RPM and 2M TPM, while Model Studio regional tables can show different ceilings. Treat rate limits as portal- and region-specific.

Multimodal work: Qwen is a better conceptual fit than text-only DeepSeek for screenshots, scanned files and video. Confirm that the chosen interface actually accepts the modality: Responses currently does not accept video.

Local deployment: The official BF16 weights, custom license and vLLM/SGLang/TokenSpeed serving guidance now exist. The raw 2.4T model remains far too large for casual workstation deployment; evaluate it only with suitable multi-node infrastructure and a license review.

Risks and evidence gaps

• Independent evidence lag: Arena offers an encouraging but narrow signal; broad production testing remains thin.
• Open-weight boundary: The checkpoint has shipped, but its custom license and different modality/context behavior mean it should not be presented as identical to the hosted product or as unconditionally open source.
• Documentation churn: Pricing, cache and output pages disagree in material ways, while QwenCloud’s seven-day linked-data deletion statement and 30-day store=true retention statement remain unreconciled.
• Long-context uncertainty: Maximum window size says nothing about accuracy decay, latency or memory cost across the full window.
• Agent reliability: Vendor benchmark averages do not reveal recovery behavior, partial failures, tool misuse or security exposure in a customer environment.
• Operational concentration: A 2.4T model can be economically attractive as an API while remaining difficult and expensive to operate independently.
• Plan-specific privacy: Training-use and retention language varies by portal and plan; generalized assurances are unsafe.

Final verdict

Use Qwen3.8-Max now if you want one hosted model for code, documents, images and video, can tolerate launch-stage documentation, and will validate it against your own acceptance tests. Its $2/$6 price makes a serious pilot easy to justify.

Wait if your decision depends on workstation-scale local deployment, permissive weight licensing, reproducible independent coding-agent results or unambiguous retention terms. Kimi K3 currently has the stronger broad independent performance record and downloadable multimodal weights. DeepSeek V4 Flash 0731 is dramatically cheaper and easier to self-host, but text-only; Artificial Analysis measured its first-party endpoint faster than Kimi’s, while a comparable Qwen measurement was not yet available.

Qwen3.8-Max has earned a place in the frontier buying conversation. It has not yet earned a blank cheque.

Qwen 3.8 Max FAQ

What is Qwen 3.8 Max?

Qwen3.8-Max is Alibaba’s production Max-tier multimodal mixture-of-experts model, launched August 3, 2026. It has 2.4T total parameters, 95B active per inference and a one-million-token context window. The production API ID is qwen3.8-max.

Is Qwen3.8-Max the same as qwen3.8-max-preview?

No. qwen3.8-max-preview debuted through Token Plan on July 19. qwen3.8-max is the August 3 production identity. Preview results should not be presented as production tests unless the behavior is reconfirmed.

Is Qwen3.8-Max open source or open weight?

Yes, the official Qwen3.8-2.4T-A95B weights shipped August 12 under the custom qwen3.8-max license. Call it open weight rather than unconditionally open source, and do not conflate the text-only open model with the hosted multimodal product.

Can Qwen3.8-Max run locally?

Yes, but only with data-centre-class infrastructure. Official BF16 weights and serving guidance now exist; the raw 2.4T pool is still roughly 1.2 TB even at a theoretical four bits per parameter, before runtime overhead.

What inputs does Qwen3.8-Max support?

Official product documentation lists text, image and video input with text output. Interface support differs: video should use Chat Completions or DashScope rather than Responses.

What is the context window and maximum output?

QwenCloud lists a one-million-token context, about 991.8K maximum input and 131.07K maximum output. Thinking mode slightly lowers maximum input. Real accuracy at the full window remains unproven independently.

How much does Qwen3.8-Max cost?

QwenCloud lists $2 per million input tokens and $6 per million output tokens, with separate cache and tool rates. Alibaba Model Studio uses region-specific CNY pricing, so buyers should specify portal, region and currency.

Is Qwen3.8-Max better than Kimi K3?

Not categorically. Qwen is cheaper and appears highly capable across multimodal and agent tasks. Kimi K3 has downloadable weights and the stronger broad independent result available at the cutoff. Benchmark both under one harness.

Is Qwen3.8-Max better than DeepSeek V4 Flash 0731?

Qwen is the stronger fit for multimodal work and has a much larger active path. DeepSeek is text-only but dramatically cheaper, MIT licensed and supported by a current independent evaluation. The better model depends on task acceptance rate, not parameter count alone.

Does Alibaba use Qwen API data for training?

There is no safe universal yes-or-no answer. General QwenCloud API guidance says prompts and completions are not used for training, while China Personal Token Plan permits content use for improvement and optimization. Responses also stores state by default. Verify the exact plan, endpoint, region and contract.

Useful Kingy.ai context

• Qwen 3.6 Max explained
• Alibaba Qwen Team profile
• Kimi K3 benchmarks, specifications and pricing
• Kimi K3 versus DeepSeek V4 Flash
• DeepSeek V4 Flash 0731 on a 128GB Mac Studio
• The state of open-weight AI models
• Kingy.ai Model Intelligence Hub