What's this about?
Alibaba released Qwen3.8-2.4T-A95B, a 2.4-trillion-parameter model with 95 billion active parameters — by parameter count, the largest model whose weights anyone has ever published for free download. The full-precision files total about 4.9 terabytes. A compressed "1-bit" version from Unsloth brings that down to 397GB, which is the number that got people excited: that fits on hardware a well-funded individual could actually buy.
What exactly are "open weights"?
When a lab releases a model's *weights*, it publishes the actual learned numbers inside the model, so you can download it and run it on your own machines forever — no API, no per-token bill, no company able to switch it off or change it under you. This is different from "open source" in the traditional sense: you usually don't get the training data or training code, just the finished brain. The comparison is being handed a working engine versus being handed the factory blueprints.
What's the community saying?
- "Is this the largest open-weight release ever?" was the thread's organizing question, and the answer settled on yes by parameter count — a milestone moment for the open ecosystem.
- The quantization math is the real story. The most-cited comment worked through Unsloth's numbers: a 1-bit quant at 397GB with 95B active puts roughly Opus-4.5-level capability on a machine a normal person could buy, at usable tokens per second.
- Serving it is genuinely hard right now. Because Alibaba only released BF16 and FP8 — no quantization-aware-trained int4 — someone with deep pockets has to produce a good quant with proper calibration data before most people can run it.
- The best features were held back. Real disappointment that vision input, non-thinking mode, built-in tools, and the 1M-token context are reserved for the closed Qwen3.8-Max; the open release is deliberately the lesser model.
- Skepticism about the model card. More than one person said the card "looks almost too good to be true," and asked whether the published benchmarks were run on BF16 or the lower-precision FP8 weights.