Thinking Machines releases Inkling 2 Mini, a looped MoE model aimed at local users
Summarized from the Hacker News discussion because the article couldn't be read
- Commenters report the release includes a looped transformer model and a varlooped research variant, which one commenter says has no other working open source equivalent.
- A commenter says the model is a MoE with active parameters in the same range as Qwen 3.6 35B A3B, and that Unsloth has already published GGUF quants.
- Quoting the company's Reddit note, a commenter says the models are trained end-to-end in nvfp4 and keep about 90% top-1 agreement even at 1-bit quantization, per Unsloth's observations.
- Commenters say the claimed parity with Qwen3.8-27B is not yet verified, and one reports that Inkling 2 Small performs better in practice than its benchmarks suggest.
Hacker News opinions
Benchmarks look good, and Inkling 2 small felt better than the benchmarks suggested. The varlooped variant is what I care about for research, since I don't know of another working open source looped model.
Unsloth already put up GGUF quants, but the llama.cpp PR for looped architecture support is still open, so you can't test it locally yet.
Inkling is the best prose I've seen from an LLM in my sampling, easy to read. For code I still reach for DeepSeek.
It's a MoE, so decode should be way faster than Qwen 3.8 27B on my 7900XTX with similar quality. Will report back.
Looped transformers might finally be good enough for real use. Rumor says OpenAI uses them too, but nobody has confirmed it.
If the benchmarks hold up that's bonkers, but I'm not betting on it until real-world use backs them.
I want these running on commodity hardware, but the GPU market has to cool down first.
A MoE with several agents, each with its own KV cache in VRAM, might be the first model that runs multi-agent reasoning on local hardware. Beating Qwen 3.8 would be a big deal.