Alibaba released Qwen 3.8-Max in early August 2026. If you run qwen3.7-max in production, the upgrade decision is practical: 3.8-Max improves agentic and research performance, accepts image input, has a lower list price, and is the first Max-class Qwen model with promised open weights.
The trade-off is current cost. Qwen 3.7-Max has a limited-time 50% promotion, so it is cheaper than 3.8-Max today. Benchmark gains also vary by workload: some are large, while GPQA Diamond is effectively unchanged.
For a full introduction to the new model, read the Qwen 3.8-Max explainer. For the previous flagship, see what Qwen 3.7 brought to the table.
Before you compare benchmarks
Every benchmark below comes from Alibaba’s official Qwen 3.8 release post. This matters because both Qwen models were evaluated in the same vendor run, making the deltas internally consistent.
However:
- Independent verification is still pending.
- Most coding rows use the Claude Code harness.
- Several benchmarks are Qwen in-house evaluations.
Use these numbers to prioritize what to test in your own workload, not as a substitute for production evaluation.
Quick comparison
| What changed | Qwen 3.7-Max | Qwen 3.8-Max |
|---|---|---|
| Terminal Bench 2.1 | 74.5 | 86.6 |
| SWE-bench Pro | 60.6 | 67.7 |
| PaperBench | 64.8 | 93.0 |
| IFBench | 79.1 | 82.8 |
| GPQA Diamond | 92.4 | 92.6 |
| HLE | 41.4 | 43.6 |
| List price, per 1M input tokens | $2.50 | $2.00 |
| List price, per 1M output tokens | $7.50 | $6.00 |
| Current promotional price | $1.25 in / $3.75 out | $2 in / $6 out |
| Image input | No | Yes |
| Disclosed architecture | Not disclosed | 2.4T total, 95B active MoE |
| Open weights | Never released | Promised “next week” (~Aug. 10) |
| Context window | 1M tokens | 1M tokens |
Benchmark deltas: prioritize agentic and research workloads
The largest gains are on long-horizon tasks that require planning, execution, and self-correction.
Terminal Bench 2.1: 74.5 → 86.6
This 12-point increase is the strongest reason to test 3.8-Max for terminal-driven agents.
In Alibaba’s table, Qwen 3.8-Max scores above Claude Opus 4.8 and Fable 5, both at 84.6 on this row. GPT-5.6 Sol remains ahead at 88.8.
Test 3.8-Max first if your agent:
- Runs shell commands
- Debugs deployment failures
- Edits repositories through terminal tools
- Performs multi-step operational work
SWE-bench Pro: 60.6 → 67.7
A 7-point gain on software engineering tasks is meaningful, but it does not make Qwen the leader in Alibaba’s comparison. Fable 5 scores 80.0 on the same row.
Use this result as a reason to benchmark Qwen 3.8-Max against your own issue-resolution tasks, not as proof that it is the best option for every coding workflow.
PaperBench: 64.8 → 93.0
This is the largest change in the table: a 28-point increase for reproducing AI research papers.
If your application handles research reproduction, long technical documents, or multi-step scientific reasoning, this is the benchmark delta most likely to affect your results.
IFBench: 79.1 → 82.8
Instruction following improves by nearly 4 points. Qwen 3.7-Max was already strong in Alibaba’s comparison, so 3.8-Max extends an existing strength.
This improvement is relevant for:
- Structured extraction
- Tool-use prompts
- Multi-constraint generation
- Workflow agents with strict output formats
GPQA Diamond: 92.4 → 92.6
This result is effectively flat.
If your workload is mostly graduate-level science Q&A or knowledge retrieval, the upgrade may not improve output quality enough to offset the current promotional price of 3.7-Max.
HLE: 41.4 → 43.6
Humanity’s Last Exam improves by 2 points, but Qwen 3.8-Max still trails frontier models in Alibaba’s table. Fable 5 scores 53.3, while GPT-5.6 Sol scores 47.2.
What the benchmark pattern means
Use Qwen 3.8-Max when your workload depends on:
- Terminal and computer-use agents
- Research workflows
- Long technical reasoning
- Stronger instruction following
- Image-based inputs
Keep Qwen 3.7-Max when your workload is primarily general Q&A, summarization, or knowledge tasks that already meet your quality bar.
For the full benchmark table and methodology details, see the Qwen 3.8 benchmarks breakdown.
Architecture: Qwen 3.8-Max is easier to capacity-plan
Qwen 3.8-Max is a 2.4 trillion parameter mixture-of-experts model with 95B active parameters per forward pass. It is built on the Qwen 3.5 architectural foundation.
The active ratio is roughly 4%, which matters for serving economics: inference compute is based on the active parameters rather than all 2.4T parameters.
Alibaba did not publish equivalent scale details for Qwen 3.7-Max. Qwen 3.8-Max has documented parameter counts and architecture information in the Model Studio documentation and release post.
For open-weight context, Kimi K3 is larger at 2.8T total parameters and 104B active parameters.
Multimodal support: image input is new
Qwen 3.7-Max is text-only. Qwen 3.8-Max accepts image input natively.
Alibaba’s multimodal table reports these notable Qwen 3.8-Max scores:
- MathVision: 95.2
- LogicVista: 91.9
- OSWorld-Verified: 86.1
There is no Qwen 3.7-Max comparison because the older model does not support images.
This can simplify applications that previously routed different inputs to separate text and vision models, including:
- Screenshot understanding
- Document-image processing
- UI automation
- Chart extraction
Alibaba also demonstrates long-document and video understanding in its release post. Treat those as demonstrated capabilities and verify supported request formats in the API documentation before designing around them.
Pricing: Qwen 3.8-Max is cheaper at list price, but not today
At list price, Qwen 3.8-Max costs:
- $2 per 1M input tokens
- $6 per 1M output tokens
Qwen 3.7-Max lists at:
- $2.50 per 1M input tokens
- $7.50 per 1M output tokens
That makes Qwen 3.8-Max 20% cheaper than its predecessor at standard pricing.
However, Qwen 3.7-Max currently has a limited-time 50% promotion:
- $1.25 per 1M input tokens
- $3.75 per 1M output tokens
At those promotional rates, 3.7-Max costs 38% less than 3.8-Max. Check the official Model Studio pricing page before setting production budgets.
Account for reasoning-token cost
Qwen 3.8-Max defaults to:
{
"reasoning_effort": "xhigh"
}
Thinking tokens are billed as output tokens. A model with a lower listed input price can still cost more per request if it generates substantial reasoning output.
Before switching, test at least three settings:
{
"reasoning_effort": "xhigh"
}
{
"reasoning_effort": "medium"
}
{
"reasoning_effort": "low"
}
Compare task success, latency, output-token usage, and total cost. For cache economics and per-task calculations, see the Qwen 3.8 pricing guide.
If Max-tier pricing does not fit your workload, qwen3.7-plus is the value option at $0.40 input and $1.60 output per 1M tokens, currently 20% off. See the Plus vs. Max comparison.
Open weights: promised, but not available yet
Qwen 3.7-Max never shipped open weights, and Alibaba never indicated that it would.
Qwen 3.8-Max is different: Alibaba’s release post says weights will be released on Hugging Face and ModelScope “next week,” which indicates roughly August 10, 2026.
As of August 3, 2026, the weights are not downloadable. Plan based on what is available today, not on a future release.
Even if weights arrive, a 2.4T parameter model is a multi-node self-hosting project, including when quantized. For most teams, the immediate effect is more likely to be third-party hosted options and pricing pressure rather than self-hosting in a local rack.
If open weights are a procurement or compliance requirement, that distinction may determine the decision:
- Qwen 3.7-Max: no open weights
- Qwen 3.8-Max: open weights promised
What does not change
Context window: 1M tokens
Both models support a 1M-token context window. Qwen 3.8-Max does not expand it, but it maintains the same capacity at a flat list price across the full context.
API migration: mostly a model ID swap
Both models run through Alibaba Cloud Model Studio using OpenAI-compatible endpoints.
For basic text requests, switch model IDs:
- "model": "qwen3.7-max"
+ "model": "qwen3.8-max"
This is not a full API migration project.
Qwen 3.8-Max also adds an Anthropic-compatible API surface. If your stack uses that protocol, exercise it with a client such as Apidog before changing production traffic.
Reasoning controls are explicit
Qwen 3.8-Max documents these controls:
reasoning_effortenable_thinkingpreserve_thinking
reasoning_effort supports xhigh by default, plus medium and low.
If you previously relied on implicit reasoning behavior from Qwen 3.7-Max, make the setting explicit in 3.8-Max and include it in your benchmark matrix.
Choose an upgrade path
Upgrade to Qwen 3.8-Max now if
Your workload is agentic, research-heavy, or image-based.
Prioritize an upgrade when you need:
- Terminal agents
- Research-paper reproduction
- Screenshot or document-image processing
- UI automation
- Better instruction following
- A path toward open weights
The Terminal Bench increase from 74.5 to 86.6 and PaperBench increase from 64.8 to 93.0 are large enough to validate against production tasks immediately.
Keep Qwen 3.7-Max during the promotion if
Your workload is mostly in the flat-performance zone:
- General Q&A
- Summarization
- General chat
- Knowledge-oriented requests
GPQA Diamond changes by only 0.2 points. If 3.7-Max already meets your quality threshold, its $1.25 / $3.75 promotional rate is currently the best per-token price among the two Max models.
Set a recurring reminder to check the promotion. If it ends, Qwen 3.8-Max at $2 / $6 becomes cheaper than Qwen 3.7-Max at its $2.50 / $7.50 list price.
Move to qwen3.7-plus if cost is the primary constraint
At $0.40 input and $1.60 output per 1M tokens, qwen3.7-plus is 5x cheaper than Qwen 3.8-Max on input tokens.
Use it for routine tasks such as:
- Classification
- Extraction
- Templated generation
- Simple transformations
- Low-risk structured outputs
Read the Qwen 3.7-Plus overview before paying for Max-tier capability by default.
Test the switch before changing production
Vendor benchmarks cannot predict behavior on your prompts, schemas, tools, and data. Run a side-by-side evaluation using representative production requests.
1. Use one endpoint and parameterize the model
Both models use the same OpenAI-compatible Model Studio API surface. Define the model ID as an environment variable:
{
"model": "{{qwen_model}}",
"messages": [
{
"role": "user",
"content": "Summarize this deployment incident and return valid JSON."
}
],
"reasoning_effort": "{{reasoning_effort}}"
}
2. Create two environments
Create environments such as:
qwen-3.7-max
qwen_model=qwen3.7-max
reasoning_effort=xhigh
qwen-3.8-max
qwen_model=qwen3.8-max
reasoning_effort=xhigh
3. Run the same test set against both
Include requests that represent actual traffic:
- Common requests
- Known failure cases
- Tool-use tasks
- Long-context prompts
- Structured-output prompts
- Image inputs, if relevant
4. Compare more than output quality
Track:
| Metric | Why it matters |
|---|---|
| Task success rate | Measures whether the answer solves the problem |
| JSON/schema validity | Detects integration failures |
| Input and output tokens | Determines actual cost |
| Thinking-token usage | Important because it bills as output |
| Latency | Affects user experience and agent loops |
| Retry rate | Often reveals hidden reliability costs |
In Apidog, point a collection at the Model Studio OpenAI-compatible endpoint, then switch environments between qwen3.7-max and qwen3.8-max. Save representative requests as test scenarios and rerun them when Alibaba updates models or changes promotional pricing.
That turns an upgrade debate into a repeatable evaluation. Download Apidog and compare both models against your own traffic before changing a production configuration.
FAQ
Is Qwen 3.8-Max cheaper than Qwen 3.7-Max?
At list price, yes:
- Qwen 3.8-Max: $2 input / $6 output per 1M tokens
- Qwen 3.7-Max: $2.50 input / $7.50 output per 1M tokens
At current promotional pricing, no. Qwen 3.7-Max costs $1.25 input / $3.75 output per 1M tokens, 38% less than Qwen 3.8-Max. The promotion has no published end date.
See the Qwen 3.8 pricing guide for the full cost breakdown.
Do I need to change code to switch from qwen3.7-max to qwen3.8-max?
For basic usage, no. Both models use the same OpenAI-compatible Model Studio endpoints, so the minimum change is:
- "model": "qwen3.7-max"
+ "model": "qwen3.8-max"
Set reasoning_effort explicitly because 3.8-Max defaults to xhigh, and thinking tokens bill as output. Image requests are a 3.8-Max-only feature and require the standard image-input message format.
Does Qwen 3.7-Max have open weights?
No. Qwen 3.7-Max has not released open weights.
Qwen 3.8-Max is the first Max-class model with promised weights, expected on Hugging Face and ModelScope around August 10, 2026. As of August 3, 2026, they are not downloadable.
Is Qwen 3.8-Max better than Claude or GPT?
It depends on the benchmark and these are Alibaba’s own results.
In Alibaba’s table, Qwen 3.8-Max:
- Beats Opus 4.8 and Fable 5 on Terminal Bench 2.1
- Leads the comparison on PaperBench and IFBench
- Trails Fable 5 on SWE-bench Pro, 67.7 versus 80.0
- Trails frontier models on HLE
Wait for third-party evaluations before making broad capability claims.
Top comments (0)