This release adds the LiquidAI LFM3 family (text, vision, audio and omni) together with a new experimental streaming-audio API in mtmd, and wires realtime audio in and out of the server through the new /v1/realtime and /tts endpoints. A batch of new architectures landed: DeepSeek V4.1 Flash, SenseNova U1, Beam, Kolibri-1, Flint, Muse Spark, Granite Switch Decision, plus decision/embedding models from LiquidAI, Baidu and NVIDIA. The core got sparse-attention (QSA/DSA) reworked onto a fused masking op, a 50/50 SLRU cache for host-resident MoE experts, and cleaner no_alloc memory handling; the server got router memory-management and tool-calling improvements; the UI can now browse and download models straight from HF Hub, search conversations and compact them. Finally, ggml was bumped to v0.28.0, bringing GGML_OP_TOP_K_MASK, -1 expert-index semantics and large backend performance work (CUDA chunked graph launches, SYCL graph replay, OpenCL Q2_K/Q3_K, Hexagon HTP rework, KleidiAI SVE kernels).
I’m interested to try the performance from this change:
MoE experts kept in host memory now use a 50/50 SLRU cache instead of a plain LRU cache llama/30637
I’m curious, are any of you running decision models locally, and if so for what?
One use case is auto mode for your local coding agent (automatic permission approval).
I’ll have to try that out
I still don’t understand their purpose. 😄
It’s in the name. If you need an LLM to answer a multiple-choice question, it just does that, without the possibility of veering off into a tangential rant about the nature of free will. These answers are also dirt cheap once the setup is provided - even compared to any normal LLM emitting a single token.
So basically if you ask “Does this e-mail look suspicious, y/n?” it cannot take ten seconds to answer “Purple.”
Alright, so here’s my Q, there’s a multiple choice question about whether something should be included in the scope of a code change. An LLM is planning the change and it asks that question. A decision model could look at that and decide A or B. Wouldn’t this model need to have wider understanding of the codebase and LLM decisions so far in order to do a good decision? Wouldn’t the LLM be in a better position to do this decision? The LLM wasn’t sure, why does the decision model’s answer be better?
Or am I misunderstanding the purpose of these?
I think you’re misunderstanding the purpose of them, to be honest. They’re meant to be used instead of LLMs instead of in conjunction with. Let’s say you have 100 emails, you have 4 tags; “important”, “normal”, “low priority” and “spam”. These decision models are made to cheaply and quickly classify those 100 emails into one of those 4 options. They tell you “this email is spam, with 83% confidence”. Just by giving it the email contents.
An LLM will also be able to determine that and think about it and probably do a better job at it. But it will be an order or magnitude slower and more expensive at that task.
Feel free to correct me if I’ve misunderstood something haha
EDIT: Let me immediately correct myself here. They can absolutely be used in conjunction with LLMs, just not in the way you’re thinking. It would rather be as a pre-sorter for something like “can this task be done by a cheap local LLM or should I pass it to opus” and it chooses between those. It isn’t “smarter” than full LLMs so isn’t deciding between code change options. It would rather be a cost/time saver in that case.
OK that latter use case makes some sense to me.
without the possibility of veering off into a tangential rant about the nature of free will.
you can do that with a regular LLM by using structured responses
It’s not novel, it’s just better. The inference is simpler and faster. Output weights are additionally supposed to be a zero-sum measure of confidence rather than a stochastic chain.
I agree, I was just pointing out that the issue you quoted earlier is easily avoided
for home users they would be more palatable if they didn’t require any extra VRAM, if it was just a different “mode” for your normal LLM
That’s one thing I like about unified memory, but yeah.
I know some people turn qwens into decision models, but to get them calibrated you still have to fine tune the weights. And then it’s not the same weights anymore.
What does calibration refer to in this context?
