Alibaba debuts Qwen3.5, a 397B-parameter open-weight multimodal AI model that it says is 60% cheaper to use and 8x better at large workloads than Qwen3
ReutersEduardo Baptista
Context & Ripple Effects
Alibaba’s Qwen line had already progressed from Qwen2.5-Max performance claims to a Qwen3 update focused on reasoning, accuracy and multilingual understanding. Qwen3.5 makes the next competitive variable explicit: multimodal and agentic capability paired with claimed lower operating cost at large workloads.
The release also sits before Alibaba’s later Qwen3.8 Max preview and stated open-weight plan, suggesting a continuing strategy of using increasingly capable Qwen models to compete across both model quality and deployment economics.
First-order effects
Developers and enterprises can evaluate an open-weight, multimodal Qwen model for visual task execution, while Alibaba positions Qwen3.5 as a lower-cost option than Qwen3 for large workloads.
Alibaba raises the performance-and-cost benchmark for its own Qwen ecosystem, with its 60% cheaper-use and 8x large-workload claims becoming the immediate basis for buyer comparisons.
Second-order effects
Competing model providers face pressure to demonstrate not just benchmark quality but comparable cost per deployed workload, particularly for multimodal and agentic use cases.
Open-weight availability can give model buyers more leverage: teams able to run or adapt models themselves have another option alongside managed proprietary APIs.
Third-order effects
If performance gains continue to arrive with materially lower serving costs, model competition will increasingly turn on cost per useful task rather than parameter counts or headline benchmarks alone.
The combination of open weights, multimodality and task execution could shift differentiation toward deployment tooling and workflow integration; whether that occurs depends on real-world reliability beyond Alibaba’s stated results.
The trend: Frontier-model competition is broadening into a contest to deliver capable multimodal agents at an operating cost enterprises can sustain.
Qwen 3.5 has been released. 397B parameters including 17B active ones. This is the flagship open sourced version. Looks to be reasonably better than Qwen3 235B. I've found larger Qwen models to not be all that great. Given the recent Minimax & GLM releases, Qwen has to also have …
The new chonky Qwen 3.5 looks pretty solid, beating their own Qwen3-Max model everywhere and is much better at vision benchmarks than Qwen3-235B-A22B-VL Now what I sadly haven't seen is anything on reasoning efficiency. [image]
So speaking of benchmarks, what can be said of the new open Qwen? First, it completely destroys Qwen3-VL-235B ofc, but more surprisingly it outscores Qwen3-Max-thinking. All the while it's the same model as “Plus”. Plus just has 1M context and some more bells and whistles. [image…
🚀 Qwen3.5-397B-A17B is here: The first open-weight model in the Qwen3.5 series. 🖼️Native multimodal. Trained for real-world agents. ✨Powered by hybrid linear attention + sparse MoE and large-scale RL environment scaling. ⚡8.6x-19.0x decoding throughput vs Qwen3-Max 🌍201 [image]
BREAKING Qwen3.5-397B-A17B weights have been uploaded to Hugging Face > 397B parameters in total > 17B active parameters per token > 256K context window > expendable to 1M token more soon once i am done running my evals [image]
The new @Alibaba_Qwen Qwen3.5-397B-A17B is live on OpenRouter now! This multimodal model uses a hybrid architecture combining linear attention with sparse MoE for higher inference efficiency. Available as both the open weights version and Qwen3.5 Plus with extended 1M context.
You can now run Qwen3.5 locally! 💜 Qwen3.5-397B-A17B is an open MoE vision reasoning LLM for agentic coding & chat. It performs on par with Gemini 3 Pro, Claude Opus 4.5 & GPT-5.2. Run 4-bit on 256GB Mac / RAM. Guide: https://unsloth.ai/... GGUF: https://huggingface.co/... [image…
Qwen3.5 is Live! Today we openweight the first model, Qwen3-397B-A17B, which is a native multimodal model supporting both thinking and non-thinking modes. We have strengthened its coding and agentic capabilities to foster productivity for developers and enterprises. Hope you
A clarification of Qwen3.5 Plus and 397B: 1. for opensource, we follow the tradition to make parameters apparent so we use the name with the number of total parameters and active params. 2. Qwen3-Plus is a hosted API version of 397B. As the model natively supports 256K tokens, [i…
Let's do the KV cache math for Qwen3.5: - KV heads: 2 - Head dimension: 256 - gated attention layers: 15 - bytes per element (BF16): 2 2 x 256 x 15 x 2 = 15 360 This is the same for K and V. So, we multiply by 2: 30 720 bytes Roughly 31 kb per token of context. Meaning at max
Qwen3.5 runs quite well in mlx-lm. Awesome that we have a frontier-level hybrid model. The context gets longer but the inference speed and memory use barely change. Here's the Q4 generating a space invaders game on an M3 Ultra. Generated 4,120 tokens at 37.6 tok/s. [video]
🎊 Kudos to the teams at @Alibaba_Qwen on the launch of Qwen 3.5 with Qwen3.5-397B-A17B. 🙌 Developers can start building today for free: https://build.nvidia.com/... Or download and customize it with NVIDIA NeMo: https://github.com/... [video]
🎉 Meet Qwen3.5-397B-A17B from @Alibaba_Qwen, 397B total params (17B active), built for real-world multimodal intelligence — day-0 support is now live in SGLang! 👁️ Unified vision-language foundation (early fusion): stronger reasoning, coding & agents ⚡ Gated DeltaNet + sparse [im…