On July 19 at WAIC 2026 in Shanghai, Alibaba’s Qwen team unveiled Qwen3.8-Max-Preview. The numbers alone would carry any launch: 2.4 trillion parameters in a sparse Mixture-of-Experts architecture, native multimodality across text, images, video, and documents, and a 1M-token context window. But the figure that matters most this week is not a parameter count. It is a promise: the official release will be open-sourced.
Less than two weeks later, on August 3, QwenWork — Alibaba’s office-productivity suite built on the new model — entered public beta. Together, the two moves mark a strategic reversal, not a routine version bump.
What Qwen3.8-Max Actually Is
- Scale: 2.4T total parameters, sparse MoE. The per-token activation count is undisclosed — a critical omission for anyone estimating deployment cost or inference speed.
- Multimodality: the first trillion-parameter-class multimodal model in the Qwen line, unified across text, image, video, and documents.
- Context: 1M tokens, carried over from Qwen3.7-Max.
- Availability: live on Alibaba’s Token Plan, Qoder, and QoderWork at a 90%-off preview discount.
- API compatibility: speaks both the OpenAI and Anthropic protocols, so existing tooling plugs in without a rewrite.
- The open-source pledge: the official version ships with open weights, alongside a Qwen3.8-27B model.
About the claim of ranking second only to Anthropic’s Fable 5: Alibaba published no benchmark table and no model card. Neither LMArena nor Artificial Analysis has scored the preview. The positioning rests entirely on internal evaluation. For calibration, the predecessor Qwen3.7-Max scored 92.4 on GPQA Diamond, 80.4 on SWE-bench Verified, and 69.7 on Terminal-Bench 2.0 — if 3.8-Max improves on that base while adding multimodality, the ranking is plausible. Plausible, however, is not verified.
The Wider Table This Hand Was Played On
The strategic weight of the release only becomes visible against three concurrent moves.
First: the Kimi K3 shockwave. One week before Qwen3.8-Max’s debut, Moonshot AI open-sourced Kimi K3 — 2.8 trillion parameters, maxed out on long-horizon tasks, frontend rendering, and agent swarms. Elon Musk called it “impressive” on X, Russ Salakhutdinov (Yang Zhilin’s PhD advisor and a CMU machine learning fixture) praised it publicly, and developers on both sides of the Pacific gave it a standing ovation.
Second: DeepSeek V4 approaching. The full-fat DeepSeek V4 is rumored to be days away, with overseas developers already gray-testing it and screenshots circulating freely.
Third: the irreversible open-source turn. Mozilla’s State of Open AI report puts the average performance gap between open and closed models at 3.3 percent — effectively parity in coding and instruction following. OpenRouter’s data shows open models’ share of processed tokens exploding from 34 percent in January 2026 to 65 percent by June, and all five of the world’s most-consumed API models are now open weights.
Against that backdrop, Alibaba’s position had become awkward. After Junyang Lin’s departure in March, the company shipped three consecutive closed releases — Qwen3.5-Omni, Wan2.7-Image, Qwen3.6-Plus — callable only through the Bailian platform API, weights unavailable. The open-first tradition that defined Qwen for two years was broken, replaced by a dual-track pattern of small models open, large models closed.
Four months was all it took for the reversal. With Kimi K3’s 2.8T open weights hitting head-on and DeepSeek V4 about to arrive the same way, staying closed would have meant voluntarily exiting the developer ecosystem — and developers, ultimately, decide whose model becomes the de facto standard. Read that way, the pledge to open-source the official release reads as a late surrender.
How the Game Plays Out From Here
Short term (1–3 months): the official Qwen3.8-Max lands and opens, setting up a direct matchup with Kimi K3 — near-identical scale (2.4T versus 2.8T), divergent architectures (sparse MoE versus conventional), settled only by real benchmarks. DeepSeek V4 enters the same window, forming a trillion-parameter club of three; with MiniMax’s rumored M3 Pro (2.5–3T, open) it could become a four-way fight.
Mid term (3–12 months): raw capability stops being the bottleneck. The contest moves to price, tooling, and ecosystem depth — Bailian’s MaaS platform versus the Kimi ecosystem versus the DeepSeek developer community. Alibaba is simultaneously betting on the application layer with QwenWork, which is the right instinct with a caveat: big-model companies building products do not have a strong track record.
Long term (1–2 years): DeepSeek founder Liang Wenfeng put it bluntly in a recent internal exchange: closed source offers no real advantage, China has too many model companies, and the field will converge. Trillion-class training costs do the culling — what likely remains is two or three open flagships plus one or two closed giants such as Anthropic and OpenAI. Alibaba’s seat at that table depends on how deep the openness runs (weights, training data, evaluation frameworks, fine-tuning toolchain — not just a checkpoint drop) and whether the commercial loop sustains itself.
What Developers Should Do
- Try it now. The preview is 90 percent off on Token Plan, entry pricing around 35 RMB (about $5) a month. For the cost of a coffee you can kick a 2.4T flagship — no reason to hesitate.
- Do not build on the preview. Activation counts undisclosed, benchmarks unpublished, pricing subject to change. Previews are for understanding, not for depending.
- Wait for the official open release. If local deployment or cost control is your goal, evaluate when Qwen3.8-27B opens. And be realistic about the flagship: even open, a 2.4T MoE needs roughly 1.2TB of VRAM at 4-bit quantization — one 8×H200 box, barely.
- Benchmark on your own data. Run Qwen3.8-Max against Kimi K3 and a gray-test DeepSeek V4 under identical conditions in your actual workload. Arena rankings belong to other people; your business metrics are yours.
- Invest in the layer above the models. Model generations now rotate every three to six months. Toolchains, evaluation frameworks, and agent orchestration layers survive those rotations — avoid hard-coding to any single model.
The Takeaway
In the week Qwen3.8-Max debuted, three things happened in Chinese AI at once: Kimi K3 went open, DeepSeek V4 lined up for release, and MiniMax’s M3 Pro sat loaded in the tube. This is not the “war of a hundred models.” It is the trillion-parameter elimination round.
The 2.4T figure Alibaba brandished at WAIC is both an entry ticket and a declaration of war — but the table now seats more than one challenger, and the real game is only starting.
