Model
Model explorer

Qwen3.8 Max

CLOSED
Alibaba · Qwen3.8 family · released Aug 3, 2026

Alibaba's new flagship, GA-launched August 3, 2026 after a July 19 preview at WAIC Shanghai — first Max-class Qwen model with a committed (but as of this writing unshipped) open-weight release. 2.4T total parameters is confirmed; ~95B active parameters, the number that actually drives inference cost, comes from third-party technical coverage rather than an official model card, since Alibaba has not yet published one. The launch comparison table is entirely vendor-run with no model card, no independent verification, and no active-parameter disclosure; multiple outlets flagged it as an unverifiable narrative claim rather than a documented launch. Only the least self-serving coding rows (SWE-bench Pro, Terminal-Bench 2.1) are recorded here — the 'best on this chart' claims (OSWorld-Verified 86.1 — claimed ahead of GPT-5.6 Sol Max 83.2, Claude Fable 5 85.0, Gemini 3.1 Pro 76.2 — and PaperBench) are deliberately omitted, which also keeps this entry below the ranking-coverage floor until independent evaluations exist.

ReasoningCodingVisionFunction callingTool useAgentic
2679.4
Elo · rank #24
Parameters
2400B
Active params
95B (MoE)
Context
1M tokens
Architecture
Sparse MoE, 2.4T total / ~95B active parameters (exact expert count undisclosed), built on the Qwen3.5 architectural foundation, 1M-token context, native text/image/video input
License
Proprietary (open-weight release announced for Qwen3.8-Max and a smaller Qwen3.8-27B; not yet published as of this ingestion)
Languages
—
API price (in/out)
$2 / $6
Modalities
text · vision · video
Benchmark results
Bar shows position within the tracked field; marker = field best
Arena EloHuman preference1483#17
best: Muse Wonder · 1513
ExploitBenchAgents28.8%#11
best: GPT-6 Astra · 100.0%
LVBenchVision81.8%#5
best: Gemini 4 Argon · 91.7%
SWE-bench ProCoding67.7%#6
best: Claude Opus 5.5 · 89.9%
SWE-bench VerifiedCoding85.6%#15
best: Claude Opus 5 · 97.0%
Terminal-Bench 2.0Coding86.6%#10
best: DeepSeek-V4.1-Flash · 90.6%
Toolathlon (Verified)Agents72.5%#12
best: GLM-5.3-Flash · 78.4%
Run it locally
Closed weights — available via API only. No local deployment.
Input / M tok
$2
Output / M tok
$6
API price $2/$6 · each benchmark row carries its own source badge (see methodology)