Qwen 3.8 Flash Next

“it generated so many more bugs, and when the quantization is lower the reasoning traces get longer”

“The modern Qwen and DeepSeek models are like someone who is merely smart, but who has an alarmingly good work ethic.”

“I have smart coding at this very second, and I am running it on 128GB of ddr4 with a 16GB gpu.”

“Qwen3.8 FN is deeply convinced that it's going to get extra credit for doing more than it is asked.”

“I've been running Qwen 3.8 Flash Next on Strix Halo with halogen-flash-server which is a pretty good experience”

“For example on programming, terminal tasks, troubleshooting etc Qwen3.8-Flash-Next is almost as good as opus 4.8.”

“Models like Qwen 27B 3.8 and Flash Next 3.8 show there is a lot of room to pack more into less parameters.”

“qwen3.8 flash next is currently running on 10gb 3080 + 64 gb ddr5 ram at respectable 40-70t/s with 1200pp at max context”

“Someone actually ran the benchmark on Qwen 3.8 Flash Next NVFP4, which is a chat model and submitted it via PR.”

“first time a model i can run myself at good speed can do real programming and is almost as good as opus”

“Qwen3.8-flash-next is an impressive evolution; it reaches the small versions of the frontier models”

“Its token gen is slightly faster than readable. You can follow along and leave it alone based on its thinking traces.”

“it is quite a bit more efficient with the number of tokens used so on longer runs it often comes out ahead in terms of wall clock time”

“I think you should try. RTX 3090 with 64GB DDR4 worked surprisingly well with lots of headroom”

“I'm running 3.8-flash-next in the largest unsloth Q8 version on a 256GB system for daily use”

“the Q4 is just plain not as good”

“Setup was easier than I expected. I had it running locally in under 20 minutes, and the speed on my RTX 5070 is genuinely impressive.”

“Known issue with this model, I recommend setting thinking to 'medium' instead of default 'xhigh'.”

“This is game changing. My R9700 32GB is now smarter and about 2x faster than using Qwen-3.8-27B.”

“In my private C coding benchmark, Strata's Qwen Flash Next Coder IQ1_M beats Luna low and runs at 125 tok/s on my 5090”

“QFN not only seemed less delusional but also made useful observations and problem solved ways around many different website access issues.”

“Yet with some tweaking and optimizing I still manage to get strata to run at 40 to 60 t/s.”

“I've tried it but so far I have not seen anything that overly impressed me in terms of accuracy, though the speed is definitely there.”

“I'm getting mixed results using the IQ3_XXS quant which supposedly matches baseline, it feels significantly degraded.”

“Now you can run a nerfed 125B parameter model on under $800 of hardware, and it beats a less-nerfed 27b model.”

“It is more useful than Qwen3.8:27b (which is already quite good) and runs faster on my 7900 XTX”

“Nvidia's NVFP4 quantisation of Flash Next and offloading the ngram table to NVMe has worked well”

“Deepseek flash / Qwen flash next et al only need ~12-24 gb of VRAM for the active experts”

“it's just moderately intelligent - if you observe the CoT while reversing, you'll find that struggles”

“Qwen Flash Next 3.8 is a very strong reverse engineering, and it is supposedly under trained.”

“you could use Qwen Flash Next at some reasonable quant, with the new SSD hack for only keeping some of the model resident in memory”

“Flash Next outperforms every other model that is available to us on the GB10 in all of my testing”

“Paying for external inference for a much larger model like qwen 3.8-flash-next Q8 with full context makes a lot more sense”

“but it produces syntax correct and grammatically correct code comments, changelogs and readme type files”

“The largest unsloth published gguf also fits and runs just fine on a cpu-only machine with 128GB RAM”

“it does seem to be a rather insane over-thinker with the biggest token use of any model”

“The new Qwen model is the most promising for one or two Strix Halo 128GB with the low number of active parameters.”

“Overall I'm impressed with it, but setting the /effort to medium is what you usually want”

“Those benchmarks look seriously impressive.. considering how small of a MoE model this is.”

“Their "next" variants are usually undercooked, but useful for the community to verify support for inference stacks.”