What's this about?
A model release subtitled "from self-scaffolding to self-improvement," which HN's local-model community engaged with seriously. Early hands-on reports were positive: someone using the 35B-A3B variant for web scraping found it on par with Qwen3.8 27B at much higher speed and a higher quantization level — q4 versus q8.
What exactly does "35B-A3B" mean, and why does it matter for local use?
It's a mixture-of-experts notation: 35 billion total parameters, but only about 3 billion active for any given token. The total number determines how much the model can store; the active number determines how much computation each token costs. For a datacenter this is mainly an efficiency win. For someone running a model at home it's the difference between possible and impossible, because you can keep the whole model in memory while only doing the work of a 3B model per token. That's why the thread's biggest complaint wasn't about Ornith at all — it was that Qwen has signalled no MoE variant for its 3.8 line, removing exactly the architecture that makes local models viable.
What's the community saying?
- The unanswered question, asked repeatedly: is the base model open weights or trained from scratch? The article doesn't say, which several found disqualifying for evaluation purposes.
- The hands-on report carried the thread: on par with Qwen3.8 27B for scraping, much faster, at a higher quant — a concrete comparison rather than a claim.
- MoE availability is the real anxiety. Local-model users treat the absence of a 35B-A3B-class Qwen 3.8 as a meaningful loss, because dense models of equivalent quality won't fit.
- Hardware-specific complaints, which is a sign of a real user base: "397 is just too big for two Sparks even at NVFP4."