Qwen3.8-2.4T-A95B debuts as 2.4 trillion parameter open model rivaling Opus 4.8 and Fable 5

A ~5TB model, that's the full BF16 weight size, absolutely massive for anyone trying to self-host.

Unsloth already has a 1bit quant at 397GB with 95B active params, that basically puts Opus 4.5 level performance on hardware a normal person could theoretically buy, though the full lossless BF16 is 4.9TB.

Comparing a 1bit quant to the full model is misleading though, this is really a model for people with a big enough Mac or 4x Strix Halo/DGX Spark setups, use the right sized model for your hardware.

Opus 4.5 level performance is already available via deepseek-v4-flash-0731, and that's much smaller, runs comfortably on 2x or 4x RTX Pro 6000 Blackwell.

Is this the largest open weight release by param count? I think it is.

No, Kimi k3 is 2.8T-A100B and shipped with QAT 4bit at ~1.5TB, so it's still smaller in disk size than this bf16/fp8-only release.

This is basically a Kimi k3 rival but a chonker, no QAT means someone with deep pockets will need to quant it down to ~1.3TB to match k3's practical size. License is similar to k3 too: free under $50M revenue/year with limits above that for coding/productivity agent services.

The local 3.8-27B model is coming Friday, that's the one that'll actually matter to most people since 3.6 was solid for local inference.

Now that they've hit frontier raw performance, I'd like Chinese labs to focus on reasoning efficiency next instead of just bigger benchmarks.

reasoning_effort is new for me, xhigh/medium/low controls plus preserve_thinking enabled by default, curious how that helps with messy OCR merging tasks I've been fighting with on 3.0/3.5/3.6.

It's unfortunate the open weight version strips vision support and caps context at 250k while Qwen3.8-Max keeps 1M context and vision.

People have gotten vision bolted onto vision-less open weights before, like DSV4 Flash and GLM-5.2, so I expect someone will do the same trick here.

Qwen3.5 was fully open and fully featured, 3.8 nerfing thinking modes and cutting context feels kind of pointless in comparison.

DeepSeek V4-Pro-0813 (1.6T-A49B) benchmarks just leaked on WeChat and it's sitting around Fable 5 level, already up on OpenRouter too.

Isn't V4-Pro-0813 actually behind Sol, Fable, ChatGPT 5.5 xhigh and Opus 5 max though? Price to performance is incredible, but it's not top of the line.

A little disappointing this shipped in BF16 only, so no QAT was done here.

I've got a RTX 5090 and 64GB RAM, can I even run Qwen3.8 or Kimi k3 locally? Feels like these models are way out of reach for a single GPU setup.

You can run their smaller 30B-or-less models fine on a 5090, the ~30B Qwen3.8 variant is apparently landing Friday for exactly that use case.