← Back to feed

Alibaba’s Qwen team dropped a surprise on July 19: Qwen3.8-Max-Preview, a 2.4-trillion-parameter model, is live today on Qwen Cloud’s Token Plan, Qoder, and QoderWork. Full open-weight release is “coming soon.”

The announcement lands four days after Moonshot AI put Kimi K3 on Arena — a 2.8-trillion-parameter open-weight model that also claimed to trail only Fable 5. The two Chinese labs are now trading scale announcements at a pace that compresses what used to be annual leaps into weeks.

What Alibaba Is Claiming

Alibaba’s framing is direct: Qwen 3.8 is “one of the most powerful models available today, second only to Fable 5.” The claim rests on internal evaluations. No third-party benchmark results from Artificial Analysis, Arena, or LiveBench have been published. That puts Qwen 3.8 in the same position Kimi K3 occupied at launch — a self-reported tier-one contender waiting for independent confirmation.

The parameter count marks a significant jump. Qwen3-Max-Preview crossed 1 trillion parameters in September 2025. Qwen 3.6 Max Preview followed at an undisclosed parameter count in April 2026 and stayed closed. Qwen 3.8 is 2.4T and, unlike 3.6, is committed to open weights.

Access Today

Qwen3.8-Max-Preview is available through Alibaba’s Token Plan subscription tiers: $6 Lite (2,500 credits per 7 days), up to $68 Pro (40,000 credits per 7 days, 6-8 concurrent agents). The plans are compatible with OpenAI and Anthropic protocol tooling, so developers can route existing harnesses at Qwen 3.8 without rebuilding.

That access structure mirrors Alibaba’s commercial pivot. Qwen spent 2024-2025 building open-source momentum — the ecosystem now counts over 200,000 derivative models and 1 billion downloads — while reserving the top models for its cloud API. The open-weight release promise signals Alibaba wants the same community buzz that made Kimi K3 a frontpage story on launch.

The Race to 2T+

The trillion-parameter open-weight model is now a Chinese market entry point, not a ceiling. Three models are now competing at that tier:

• Kimi K3: 2.8T, open-weight, on Arena since July 16 (LiveBench 78.5, AA Intelligence Index 55+, $3/$15 per million tokens)
• Qwen 3.8: 2.4T, preview available, open weights pending, no published benchmarks
• DeepSeek V4 Pro: 1.6T, open-weight (MIT), 80.6% SWE-bench Verified, on market since April

Frontier labs in the US are operating at similar scale — GPT-5.6 Sol, Fable 5, Grok V9 — but keep weights closed. The Chinese labs are writing the open-source playbook at the frontier.

What to Watch

Independent benchmark runs will determine whether Qwen 3.8’s claims hold. Artificial Analysis typically posts within a week of a model going live on API. Arena usually adds models once they’re publicly accessible. The open-weight release itself, whenever it arrives, will trigger Hugging Face download charts that either confirm or deflate the launch hype.

Qwen 3.7 Max Preview ranked 13th globally on Arena text with a similar set of claims. Whether 3.8 closes the gap to Fable 5 or lands in the same tier as Kimi K3 — just outside the true frontier — will define how this announcement is remembered.