Alibaba Qwen logo

Alibaba Qwen: Qwen3.8 Max Preview

Preview Qwen flagship for million-token multimodal reasoning and long-horizon agentic workflows

Source-linked qwen Closed weights Released 2026-07-19
API record Report
InputT
OutputT
Input price$2/M
Output price$6/M
Context1M
Max output131.072K
Providers6
Inference availability

Providers

Provider-specific identifiers, limits, and listed prices per million tokens. Every row links back to the provider's own documentation.

Report a price
Providers offering Qwen3.8 Max Preview
ProviderProvider model IDContextMax outputInputOutputCache readCapabilitiesDocs
Alibaba Token Plan qwen3.8-max-preview 1M 131.072K ReasoningToolsJSON Docs ↗
Alibaba Token Plan (China) qwen3.8-max-preview 1M 131.072K ReasoningToolsJSON Docs ↗
NanoGPT qwen3.8-max-preview 991K 64K $1.5 $5 $0.15 ReasoningToolsJSON Docs ↗
DevPass (LLM Gateway) qwen3.8-max 1M 1M $2 $6 $0.25 ReasoningTools Docs ↗
Charm Hyper qwen3.8-max 1M 65.536K $2 $6 $0.25 ReasoningTools Docs ↗
Impossibl qwen/qwen3.8-max-preview 1M 131.072K $2.5 $7.5 ReasoningTools Docs ↗

Capability badges appear only where the provider catalog explicitly lists support. A blank cell means the source is silent, not that the feature is absent.

Listed rates

Price across providers

Input price per million tokens as published by each provider. Bars are drawn from listed rates only — no traffic weighting, since the catalog observes no requests.

Lowest input $1.5/M

Across 4 priced providers

Median input $2/M

Midpoint of listed rates

Highest input $2.5/M

1.7× the lowest listed rate

Output range $5 – $7.5

Per million output tokens

NanoGPT $1.5/MLowest
DevPass (LLM Gateway) $2/M
Charm Hyper $2/M
Impossibl $2.5/M
Cost calculator

Estimate a workload

$0.00
Excludes taxes, non-token charges, and tiered discounts.

Context limits also differ by provider, from 991K to 1M tokens. Compare the provider table above before choosing on price alone.

Specification

Capabilities

Recorded from the source catalog and provider listings.

Reasoning Yes
Tool calling Yes
? Structured output Unknown
Attachments Yes
Vision input Yes
× Open weights No
Model family
qwen
Knowledge cutoff
Not documented
License
Not documented
Release date
2026-07-19
Model ID
alibaba/qwen3.8-max-preview
Published evaluations

Benchmarks

Every result stays attached to its source, version, metric and harness. Scores from different versions are never merged, and the profile below plots each benchmark against its own population rather than on a shared scale.

Benchmark registry
Benchmark profile
Humanity's Last…Humanity's Last Exam GPQA DiamondGPQA Diamond OSWorld-VerifiedOSWorld-Verified MMMU ProMMMU Pro This model — Humanity's Last Exam: 43.6 accuracy (27.5th percentile) This model — GPQA Diamond: 92.6 accuracy (52.8th percentile) This model — OSWorld-Verified: 86.1 success rate (97.1th percentile) This model — MMMU Pro: 82.3 accuracy (75.0th percentile)

Each axis is this model's percentile among the 4 benchmarks it has published results for, measured against every other model with a score on that same benchmark. Percentiles are used because benchmarks do not share a scale — a 60 on one is not a 60 on another. Hover any point for the raw score.

Humanity's Last Exam accuracy · 2026-08-03 Source ↗
Humanity's Last Exam #15 of 20
GPQA Diamond accuracy · 2026-08-03 Source ↗
GPQA Diamond #9 of 18
OSWorld-Verified success rate · 2026-08-03 Source ↗
OSWorld-Verified #1 of 17
MMMU Pro accuracy · 2026-08-03 Source ↗
MMMU Pro #3 of 10
SWE-Bench Pro resolve rate · Claude Code · 2026-08-03 Source ↗
SWE-Bench Pro #1 of 3
Terminal-Bench — 2.1 accuracy · 2026-08-03 Source ↗
86.6 2 models scored — too few for a distribution
AutomationBench pass@1 · 2026-08-03 Source ↗
27.3 1 model scored — too few for a distribution
DeepSWE — 1.1 resolve rate · Claude Code · 2026-08-03 Source ↗
56.6 1 model scored — too few for a distribution
FrontierSWE dominance score · Claude Code · 2026-08-03 Source ↗
73.5 1 model scored — too few for a distribution
IFBench score · 2026-08-03 Source ↗
82.8 1 model scored — too few for a distribution
MLS-Bench-Lite score · Claude Code · 2026-08-03 Source ↗
41.0 1 model scored — too few for a distribution
NL2Repo resolve rate · Claude Code · 2026-08-03 Source ↗
55.9 1 model scored — too few for a distribution
Toolathlon Verified pass@1 · 2026-08-03 Source ↗
72.5 1 model scored — too few for a distribution
WideSearch F1 · 2026-08-03 Source ↗
81.9 1 model scored — too few for a distribution

Each strip shows every published score for that benchmark, with this model marked. The lighter shape behind the ticks is the density of results, and the dashed line is the median.

Cost against capability

Price and performance

Listed input price plotted against Humanity's Last Exam, the benchmark with the widest published coverage that this model appears in.

0 34 69 $1 $10 Qwen3.7 Max — $2.5/M, score 41.4 Qwen3.8 Max Preview — $2/M, score 43.6 Claude Fable 5 — $10/M, score 64.5 Claude Opus 4.7 — $5/M, score 54.7 Claude Opus 4.8 — $5/M, score 57.9 Claude Opus 5 — $5/M, score 64.7 Claude Sonnet 4.6 — $3/M, score 46.8 Gemini 3.1 Pro Preview — $2/M, score 44.4 Gemini 3.5 Flash — $1.5/M, score 40.2 Muse Spark 1.1 — $1.25/M, score 62.1 Nemotron 3 Ultra 550B A55B — $0.5/M, score 37.4 GPT-5.4 — $2.5/M, score 52.1 GPT-5.4 mini — $0.75/M, score 28.2 GPT-5.4 nano — $0.2/M, score 24.3 GPT-5.4 Pro — $30/M, score 58.7 GPT-5.5 — $5/M, score 52.2 GPT-5.5 Pro — $30/M, score 57.2 GPT-6 Astra — $10/M, score 57.2 Step 3.7 Flash — $0.185/M, score 47.2 GLM-5.2 — $1.4/M, score 54.7 Qwen3.8 Max Preview Input price per million tokens (log scale) Accuracy

The stepped line is the efficient frontier: at each price, the best score available for that money or less. A model sitting on it is not being beaten by anything cheaper. Price is log-scaled because listed rates span four orders of magnitude. Only models with both a listed price and a score on this benchmark can appear.

Catalog activity

Change log

Field-level changes detected between successful source imports.

Full change log
Price Completion5.4461 → 6.0
Price Prompt1.815 → 2.0
Price Completion5.0 → 5.4461
Price Prompt1.5 → 1.815
Price Completion7.5 → 5.0
Price Prompt2.5 → 1.5
Provenance

Sources & verification

Every figure on this page traces back to one of these records.

Methodology
Public API

Use this record

Fetch the complete source-linked model record. No key, no account, no rate-limited tier.

API documentation
Endpoint
GET https://model.kyssta.lol/api/v1/models/alibaba/qwen3.8-max-preview
curl
curl "https://model.kyssta.lol/api/v1/models/alibaba/qwen3.8-max-preview"
Common questions

Frequently asked questions

Answered directly from the stored record — nothing here is generated beyond the catalog's own fields.

What is Qwen3.8 Max Preview?

Preview Qwen flagship for million-token multimodal reasoning and long-horizon agentic workflows. It is published by Alibaba Qwen and catalogued here from Models.dev.

How much does Qwen3.8 Max Preview cost?

Listed input pricing starts at $1.5 per million tokens from NanoGPT, rising to $2.5 across 4 listed providers.

What is the context length of Qwen3.8 Max Preview?

Qwen3.8 Max Preview accepts up to 1M tokens of context and returns up to 131.072K output tokens.

Does Qwen3.8 Max Preview support tool calling and structured output?

Provider catalogs list support for tool calling, reasoning, and image input.

Which providers serve Qwen3.8 Max Preview?

6 providers list this model: Alibaba Token Plan, Alibaba Token Plan (China), NanoGPT, DevPass (LLM Gateway), Charm Hyper, Impossibl.

Are the weights for Qwen3.8 Max Preview open?

No. This model is served through hosted APIs only.

When was Qwen3.8 Max Preview released?

The catalog records a release date of 2026-07-19, last verified Sep 11, 2026.

More models from Alibaba Qwen

View all →