This week, a post went viral on Reddit's local AI community. One user said Alibaba's Qwen 3.8 27B model has excellent quality — but running "high-intensity thinking" mode on their M1 Max laptop (one of Apple's current top-tier chips) takes all night per task, calling it "completely unrealistic."

They're asking for a version that doesn't exist yet: Qwen 3.8 35B A3B. "A3B" means "3B activated" — i.e., a MoE (Mixture of Experts) architecture: 35B total parameters (medium-large), but only 3B activated per inference, making it much faster — "a bit dumber but actually usable."

This might look like a niche technical discussion, but it hits a real pain point: anyone running large models locally is searching for a size that "just barely runs." We noticed that in the post's comment section, some users were calculating VRAM, others were waiting for Nvidia's next-gen GPUs, and some had already given up — directly subscribing to ChatGPT instead.

What this is

Qwen is Alibaba DAMO Academy's open-source large model family and one of the most-downloaded Chinese open-source models globally right now. "27B" means the model has 27 billion parameters; "xhigh thinking mode" means the model runs long-chain reasoning before answering, similar to the "slow thinking" paths of OpenAI o1 and DeepSeek R1.

The 35B A3B the user wants is essentially a MoE architecture trade-off — grow total parameters to preserve capability, shrink active parameters to preserve speed. This has been the common direction for mainstream open-source models since 2025: Mistral's Mixtral and DeepSeek V3 both follow this approach.

Industry view

On the positive side, MoE is indeed the optimal solution for local large models right now. The Qwen family is rated by many local users on Reddit as "the strongest open-source model that runs at home," with a mature ecosystem and complete toolchain. If an "A3B" small-activation variant is actually released, it would further lower the barrier to local AI.

But we also need to acknowledge several counterarguments. First, A3B is not a universal fix — smaller active parameters mean limited "simultaneous thinking" capacity; long-chain reasoning (multi-step logic, complex planning) may still struggle. Second, the hardware wall is harder than expected: M1 Max and high-end RTX GPUs cost upwards of $1,000 (often several thousand), already shutting out 90% of ordinary users; heat, memory, and electricity are all pitfalls. Third, and most practically — Anthropic, OpenAI, and DeepSeek API prices have fallen an order of magnitude over the past two years, and the "cost-performance advantage" of local deployment is being rapidly eroded by the cloud.

Impact on regular people

For enterprise IT: Want to deploy AI on your company's intranet and keep data in-house? Hardware investment is higher than expected, and you need a dedicated team for model optimization and operations — it's not as simple as installing software.

For working professionals: Want to run a decent AI assistant on your own laptop to replace ChatGPT? The 2026 reality is — either spend $2,000–3,000 on high-end hardware, or keep paying monthly for cloud services. There's no "free lunch" shortcut.

For the consumer market: Chinese vendors (Alibaba, DeepSeek, Moonshot) are giving away downloadable large models for free — that's good news; but between "downloadable" and "actually runnable" lies a hardware wall. The best short-term option for ordinary consumers is likely still a subscription service.