From Unsloth’s post on Reddit:
Hey guys Thinking Machines released Inkling 2 Mini, the strongest 26.4B model (regardless if MoE or dense) that runs on ~16GB RAM.
Inkling 2 Mini is Thinking Machines Lab’s new 26.4B MoE (5.4B active) model for coding, agentic & vision workloads.With a Looped Transformers architecture and 1M context, it scores 48.2% on DeepSWE v1.1 and 77.9% on SWEBench Pro v2 (Hard), the highest for its size. Thanks TML for day 0 access.
Run on Unsloth Desktop or llama.cpp.
But looks like the model is not available in llama.cpp yet:
But if benchmarks are true, it’s somewhat on par with Qwen 3.8 27B!
Wonder how much vram it consumes - if the same trick as with qwen flash next about offloading most to ram works here.
It probably works in 8GB VRAM with the looped experts cached. I’ve seen people on Reddit with 12GB VRAM getting good results.
I’ve got it running with the latest version of llama.cpp compiled from source + my own Q6_K quant. The safety is cranked up to the point that it’s almost useless, and it’s disappointingly sluggish for a MoE model of this size – at least with the settings I’ve tried so far (which were pretty good for Qwen and gemma4).
I’m curious to see what comes of this after the abliterators get acquainted with it and the llama.cpp developers have a bit more time to poke at it… but so far it’s very meh.
I heard the same thing from other people online. I think that’s why Alibaba began releasing ‘NEXT’ Qwen models so that the community could take time to work out the kinks.
What are you guys doing to trigger the safety rails? 😅
a moe model designed for 16gb ram? christmas came early! 🎁
woa it’s even beating dense models, no way it’s actually that good
once those llama.cpp pull requests are merged I’ll try it out
I wonder how the looped layers interact with --moe-cache-mib
One of the PRs is merged now, although not the unsloth one so personally I am trying the ggml-org weights first to be on the safer side. With a novel architecture and looping layers I wouldn’t rule out a few days of problems to iron out in the implementation though…
According to reports it’s not very stable in llama.cpp right now. Unsloth fixed it in their fork, so yeah, I’d wait a little.
Ooh! Qwen4 will need to knock it out of the park, it’s nice to have competition keeping things moving. :)
Looks promising
This is what happens when Qwen strings you along for weeks with delivering anything or even concretely stating the scope of the models they’ll release.




