Qwen 3.8-Flash-Next just dropped. Can I even run it?

My phone would not stop this morning. Qwen 3.8-Flash-Next is out, the reaction videos were already up before I had finished my coffee, and none of them answered the only question I actually had: can the machine on my desk run this thing, and is it worth trying? So before benchmarking anything I did the boring version. I opened the Unsloth guide and the official Qwen model card, read them on camera, and worked out the answer from the published numbers.

This is a first look, not a deep dive. Nothing here was measured by me yet. Every number is Qwen's or Unsloth's, and I say so wherever it matters.

What it is

Qwen 3.8-Flash-Next is a 125 billion parameter mixture-of-experts model, built on a new Qwen4 architecture, with a 262K context window. The number that surprised me on the model card is the second one: only 6 billion parameters are active per token. On top of that sit 51 billion parameters of n-gram embeddings, which as far as I can tell are pre-computed lookup tables keyed on short runs of tokens (I said "do not quote me on that" on camera, and I meant it; that is the follow-up video). It also has an MTP layer built in, the multi-token prediction head I spent a whole video measuring last week on the 27B.

The comparison Qwen puts on the card is against their own 3.8-27B, DeepSeek V4 Flash, and Claude Opus 4.6 (Max), and the Unsloth page opens with "outperforms Claude-4.6-Opus (Max)". These are vendor benchmarks. I read them, I do not repeat them as fact. What I will say is that beating the 27B is expected, because even with fewer active parameters the base model is almost five times larger.

The sizes, which is what I actually came for

Build	Memory needed	Unsloth's KL divergence table
1-bit (UD-IQ1_S / IQ1_M)	75 GB	mean KLD 0.30 to 0.38, top-1 accuracy about 80 to 82%
2-bit	79 GB	0.21, 85%
3-bit	90 GB	0.10, 90%
4-bit (UD-Q4_K_XL)	112 GB	0.045, 93.5%
BF16	355 GB	the reference

Unsloth's line is that the smallest quant works on 75 GB, so you should have a 96 GB machine. That is RAM plus VRAM, or unified memory. The 1-bit build is 79% smaller than BF16 and, by their measurement, keeps about 80% top-1 accuracy.

KL divergence, if the term is new to you: it measures how far a quantized model's next-token probabilities drift from the full-precision model's. Lower is closer. Top-1 accuracy is how often the shrunk model picks the same most-likely next token as the original. So the 4-bit agrees with BF16 on the top token about 93 times in 100, the 1-bit about 80. That is Unsloth's own method, not a standard benchmark, and it says nothing about speed.

One detail worth knowing before you download: the n-gram and per-layer-embedding tables are kept at 4-bit in every build, including the 1-bit one. They are lookup tables with a random access pattern, so quantizing them hard damages the model, and it is also why the 1-bit and 2-bit files are almost the same size. The savings come from the dense layers, not the tables.

So, can I run it?

Technically, yes. My mini PC has 128 GB of unified memory, so the 75 GB 1-bit fits with room for the OS and whatever else is running, and the 2-bit at 79 GB probably does too. Whether that is worth doing is a different question. The 4-bit, the usual sweet spot for a model this size, is 112 GB, which on my box is not happening.

The hopeful part is the 6 billion active parameters. On the 27B with MTP at depth 2 I get around 10 tokens per second depending on the task. If only 6B of this model runs per token, there is a real chance it is faster than the 27B despite being much bigger on disk. That is the thing I want to see, and I have no number for it yet.

The less hopeful part is the price of admission. 96 GB is multiple GPUs or a top-spec unified-memory machine, and the new Macs with the throughput to make a model like this pleasant cost about what a car does. This is the same feeling I had with DeepSeek V4 Flash: the models keep getting better and harder to run at home in the same breath. Sampling settings, at least, match the 27B exactly (temperature 1.0, top-p 0.95, top-k 20 in thinking mode), so my existing benchmark configs carry over.

I checked Artificial Analysis for independent numbers and there is nothing listed yet. I will come back to that.

What happens next

Benchmarks on my Strix Halo rig, the same One-Shot suite I ran on the 27B: tokens per second, memory headroom, and whether 6B active means anything on unified memory. That is the follow-up. If you are already running Qwen 3.8-Flash-Next, especially on unified memory, tell me your quant, your memory and your tokens per second in the comments; I want to compare.

Sources