Tencent open-sourced Hy4 preview on August 28, a 770-billion-parameter MoE model with 49 billion activated parameters and a context window exceeding 1 million tokens1,5,6. Two days earlier, Alibaba released Qwen3.8-Flash-Next, a 125-billion-parameter multimodal MoE model it described as "an early preview of the architecture used in Qwen4"2,4. ANALYSIS The near-simultaneous drops mark a new phase in which China's largest platform companies are racing not just to build frontier models but to open-source them, compressing competitive cycles and raising the bar for every open-weight alternative.
Why it matters
Open-weight models from Chinese labs have moved from curiosities to credible alternatives for production workloads. Hy4 preview is priced at $0.834 per million input tokens and $2.501 per million output tokens, with cache hits at $0.042 per million tokens. WorkBuddy and CodeBuddy will offer free access for a two-week launch window, while free access to the prior Hy3 model has been extended through September 30. ◆ Tencent is using aggressive pricing and free trials to seed adoption across its own product surface and third-party platforms, a distribution playbook that ties the model's success to Tencent's broader cloud and productivity ecosystem.
The big picture
Tencent's internal blind evaluation, involving 163 experts and 203 engineering tasks, gave Hy4 preview an average score of 2.99 out of 4, compared with 2.92 for GLM 5.3 and 2.94 for Kimi K3. Tencent also said the model helped optimize parts of its own training and inference systems, increasing end-to-end throughput by 31.8% against a baseline. The model is available through Tencent Cloud TokenHub and OpenRouter.
Alibaba's Qwen3.8-Flash-Next, meanwhile, takes a different tack. Rather than competing on raw parameter count, it previews architectural changes headed for Qwen4. Alibaba confirmed in its release blog that Qwen3.8-Flash-Next is "a multimodal MoE model that also serves as an early preview of the architecture used in Qwen4". The company stated that Qwen3.8-Flash plays the same role that Qwen3-Next played for Qwen3.5, introducing developers to a hybrid Gated DeltaNet and Gated Attention design. With 125 billion total parameters but only 6 billion active, the model targets a balance of capability, latency, and cost.
ANALYSIS The two releases reflect distinct strategic logics. Tencent is pushing scale (770 billion parameters, million-token context) and tying the model to its own productivity tools. Alibaba is seeding architectural familiarity so developers begin building on Qwen4's foundations before the full model family ships.
Between the lines
The practitioner community is already stress-testing both lineages. One Reddit user reported running a quantized Qwen 3.8 27B variant on a 12GB GPU with up to 200K tokens of context, calling it "more capable than Sonnet 4.6"7. Another benchmarked Qwen3.8-Flash-Next on an M4 Max system and reported it was "the first model for the year that was able to break through 94%" on a personal coding-and-knowledge benchmark8. Simon Willison noted he was exploring the model on a DGX Spark using Unsloth quantized weights.
ANALYSIS The speed at which quantized variants reach local-inference toolchains, as seen in the Reddit and DGX Spark experiments, suggests that open-weight releases from Chinese labs now plug into the same hobbyist and enterprise pipelines that once centered on a single dominant open-weight family. The competitive pressure is no longer just between labs; it flows downstream to quantization teams, inference engines, and application developers who can swap model backends in hours.
Tencent's claim that Hy4 preview "autonomously analyzed bottlenecks in its inference system and carried out multiple rounds of optimization" points to an early recursive self-improvement loop. ◆ If reproducible outside Tencent's own infrastructure, that capability would represent a qualitative shift in how open-weight models are deployed: not just as static artifacts but as participants in their own optimization.
What's next
Tencent has indicated the next batch of models in the Hy4 series is expected to roll out soon. Alibaba's explicit framing of Qwen3.8-Flash-Next as a Qwen4 architectural preview means a full Qwen4 model family is in preparation. The two-week free-access window for Hy4 preview on WorkBuddy and CodeBuddy closes in mid-September, at which point Tencent's conversion-to-paid metrics will offer the first real signal of enterprise pull.