Alibaba Officially Launches Qwen3.8: 2.4T Parameter Model Now #2 on Arena, Open-Source Coming Next Week

August 3, 2026 — Alibaba today officially launched Qwen3.8, its newest flagship AI model, putting hard benchmark numbers behind the claims it previewed at WAIC last month. The model lands at #2 on the Arena leaderboard, trailing only Anthropic's Claude series, and brings a 2.4-trillion-parameter sparse MoE architecture with 95 billion active parameters per token.

---

The Numbers That Matter

Here are the verified benchmark scores Alibaba published today:

Benchmark	Qwen3.8-Max Score	Context
PaperBench (coding agent)	93.0	+28.2 points over predecessor — a new record
GPQA Diamond (science reasoning)	92.6	Frontier-tier
OSWorld-Verified (computer use)	86.1	#1 among mainstream models
IF Bench (instruction following)	82.8	Top tier
BabyVision (visual reasoning, no tools)	82.0	Nearly 2× some mainstream models
WideSearch (general agent)	81.9	Frontier
Agent's Last Exam (general agent)	52.4	Frontier

On Arena specifically, Qwen3.8 now sits behind only Anthropic's Claude series — above GPT-5.5, above Kimi K3, above Grok 4.5. On CodeArena, it ranks #4 globally. On Vision Arena, it ranks #2 worldwide.

---

Architecture: Sparse MoE at Scale

Qwen3.8 uses a sparse Mixture-of-Experts architecture with:

• 2.4 trillion total parameters
• 95 billion active parameters per token
• 1M token context window
• Native multimodal: text, images, video, documents
• Hybrid attention mechanism jointly optimized with the MoE routing

This is the first time Alibaba has disclosed active parameters for a Qwen flagship — 95B active puts it in a competitive efficiency range versus dense models of comparable capability.

---

Coding: From Empty Folder to Production

The most striking demo: Alibaba had Qwen3.8 start from an empty folder and autonomously build a real project over 16 days without human intervention. The result is "oh-my-cli" — a self-evolving agent framework at the Hermes Agent level, now fully open-source on GitHub.

The model was given a single prompt: "Create a self-evolving intelligent agent Harness." It then:

1. Built a Loop Engineering framework from scratch
2. Orchestrated sub-agents to autonomously pick up and execute tasks
3. Accepted new requirements mid-process via DingTalk
4. Completed a production-grade deliverable after 16 days of continuous autonomous coding

The entire development process — every commit, every decision — is publicly viewable on GitHub.

---

Cowork: Professional Tasks at Scale

Alibaba demonstrated Qwen3.8 on real professional workloads:

• Legal: Reviewed hundreds of legal documents, annotated thousands of clauses — a week's work for a paralegal team — in under one hour
• Sports Analytics: Analyzed 160+ hours of basketball footage (8,400+ plays), producing per-player tactical profiles and coaching reports in minutes
• Quantitative Finance: Autonomously researched, built, backtested, and iteratively improved a complete ETF rotation strategy over several hours, reaching profitability

In digital chip design — a high-precision, physics-constrained domain — Qwen3.8 demonstrated system-level autonomous planning and closed-loop adaptive learning through thousands of iterations.

---

Pricing & Availability

API Pricing (effective today)

	Domestic (China)	International
Input	¥12 / 1M tokens	~40% of Claude Opus 5
Output	¥36 / 1M tokens	~24% of Claude Opus 5
Cache Hit	¥1.5 / 1M tokens	—

What's Available Now

• Qwen3.8-Max API via 千问AI (Qianwen AI) platform
• 千问办公 (Qianwen Office) — Alibaba's new AI agent product, powered by Qwen3.8

Open Source — Coming Next Week

• Qwen3.8-Max: full open-weight release planned for next week
• Qwen3.8-27B: a smaller 27B dense model also going open-weight

This breaks Alibaba's previous pattern — flagship "Max" models were historically API-only. If Qwen3.8-Max ships open-weight as promised, it would be the most capable open-weight model available.

---

The Competitive Landscape

Qwen3.8 enters a crowded field:

Model	Params	Open?	Arena Position
Claude Fable 5	Undisclosed	Closed	#1
Qwen3.8-Max	2.4T (95B active)	Open next week	#2
Kimi K3	2.8T (MoE)	Open-weight	Below Qwen3.8
GPT-5.5	Undisclosed	Closed	Below Qwen3.8
Grok 4.5	Undisclosed	Closed	Below Qwen3.8

The key takeaway: Qwen3.8 achieves near-frontier performance at significantly lower cost than Western alternatives (40% of Opus 5 input pricing internationally), with open weights coming next week. For production teams weighing cost vs. capability, this changes the equation.

---

What This Means

Three things stand out:

1. The pricing gap is real. ¥12 input / ¥36 output per million tokens, with international pricing at a fraction of Anthropic's rates, makes frontier-level AI accessible at a new price point — especially for Asian markets.

2. Open-weight is the new battleground. Alibaba reversing course on its flagship model (from proprietary to open-weight) signals that open release is now a competitive advantage, not a concession. Between Kimi K3's 2.8T open-weight release and Qwen3.8-Max next week, the open-weight frontier has never been stronger.

3. Autonomous coding is crossing a threshold. 16 days of unsupervised development producing a real, open-sourced framework is not a demo — it's a working system. The implications for software engineering productivity are hard to overstate.

---

The Bottom Line

Qwen3.8 is the real deal. The July preview raised eyebrows with bold claims and no benchmarks; today's launch delivers the receipts. With Arena #2 ranking, disclosed 95B active parameters, competitive pricing, and open-weight release imminent, this is the most significant Chinese AI launch of 2026 so far.

Have thoughts on Qwen3.8? The open-weight release next week will tell us even more. Stay tuned.

---

Sources: TechWeb (Aug 3, 2026), Alibaba Qwen Team official announcement, Yotta Labs, FelloAI, Bloomberg, Wan 2.7, community CodeArena tracking