2 Updates: DwarfStar and Redis new sorted sets: 1) The new version of DwarfStar is much faster in prefill both on the M5 Max and the DGX Spark. It also integrates 14 pull requests / issues fixes and other stuff. Time to upgrade :)

Now part of the DwarfStar README

Read those HN comments as it is sociologically super interesting to see people that can't cope with AI results news.ycombinator.com/item?id=4913...

Btw now that I can run MXFP4 locally and generate greedy continuations, I can tell you: not all the OpenRouter providers for DS4 flash are sane... Not going to do names, but some provider gives you the real shit, other will not.

DwarfStar branchk "ds4f-mxfp4" now can run the lossless MXFP4 DeepSeek v4 Flash GGUF I published on my Hugging Face account. It rocks even with SSD streaming in 128GB systems at > 20 t/s in case you want to try the *actual* DS4F weights released without any quantization.

I'm going to sleep but you will see the new GGUF files for DeepSeek v4 Flash here as they upload during the next hours: huggingface.co/antirez/deep... (not tested, just smoke tested. Tomorrow will be testing time).

antirez/deepseek-v4-gguf at main

We’re on a journey to advance and democratize artificial intelligence through open source and open science.

huggingface.co

I'm starting the conversion work from new DeepSeek v4 Flash checkpoint to GGUF. If the model is as good as it looks, I'll probably remove the GGLM 5.2 support from the system, since now we have a model that is best suited for local inference that is smaller. Feedbacks?

In this video of mine, while I know it is likely an utopia, I state that the world should seek in United States, Cina, Europe and the world at large to join forces to create a unique world-wide AI effort, outside of national interests. www.youtube.com/watch?v=OJEP...

Un'altra AI è possibile

YouTube video by Salvatore Sanfilippo

youtube.com

Now with the upcoming release of the new DS4F checkpoint, and I believe later with Laguna S2.2 or a new S2.1 checkpoint, DwarfStar is going to be better at zero development costs. It is important to invest in the right models that get better and are locally runnable.

Fable - Sol loops found another 23% memory reduction in the already very optimized sorted set PR. Now, let's explore radically different approaches compared to the zkiplist+hash approach.

I believe having to split large model files is an anti-pattern, and could be definitely abstracted away by Hugging Face. I look forward to changes in the way we upload and download large models, as models are headed to be ... larger :D

I reviewed a Redis me optimization and was about to submit the PR. Then I said myself: "wait, we are in 2026" and wrote a script that calls Sol and Fable at each turn, challenging to improve the work of the previous one... Soon to be continued with the results.

The real AI risk is inside the labs (a reply to Amodei's post on open weight models): antirez.com/news/172

I tried to adapt the transcript of my video "Being Linus Torvalds" into an english language blog post here: antirez.com/news/171

Warning my friends: for the sake of saying that distillation is a good thing, and not a bad one, you are falling in the trap of admitting that frontier Chinese models are *mainly* the result of distillation (which is not just not true, but also not possible).

I just published mixed q2/q3 quants of Laguna S2.1 that work in MacBook systems with 64GB of RAM with quality not far from q4. You need DwarfStar laguna-s2.1 brach for them to work, or llama.cpp. With DwarfStar the speed is 65 t/s generation, ~570 t/s prefill.

Update on the DwarfStar laguna-s2.1 branch, the GGUF files got updated, now the support for the file with Q8 projections and smaller routed experts (68GB total IIRC) is implemented as well. Speed improved significantly to 60 t/s generation, 550 t/s prefill.

I have a modest proposal: let mathematicians be the ones to crack the open problems that can be addressed with LLMs. They worked for a long time in this field. It should be their prerogative to do so.

Now the DwarfStar laguna-s2.1 branch contains a GPT5.6-Sol coded (and manually tested for apparent sanity with non trivial tasks) Laguna S2.1 implementation. 50 t/s generation, 500 t/s prefill on M5 Max. Only Metal for now. GPT5.6 Sol can write for you the other backends.

So I tested Laguna S2.1 and the results are: it is a strong coding model for its size, in certain coding tasks it performs better than DeepSeek v4 Flash (but we are waiting for an update). It is NOT a general purpose LLM, very specialized for coding apparently.

Btw if you think the argument against Chinese models is too weak to ban them internally, in the US, you should check the status of Heath Care and costs for American citizens of things that cost 1/10 elsewhere in the world. They can do it, especially with this administration.

Laguna S2.1 is, among the other things, completely unable to write correct Italian. Something that even much smaller models can do very well. I understand specialization but this is a red flag. I saw this with Q4 quants, verified with the official API (via openrouter).

This "they distilled our models!" is starting to resemble the moment when Redis was winning in the database space really hard, and certain actors started with "but it is not linearizable!", and also attacked me personally. Eventually, a few started selling Redis, before closing.

[blog post] Not just development, distribution of software may change as well: antirez.com/news/170

With DwarfStar I could have a "main" branch that is human curated, like now, and other branches that are just experimental and mostly coded via AI looping to implement new stuff, in order for users to have access/test new ideas. For Laguna S.1 this could work.

In Wohpe (Laurana, 2022) an engineer erroneously leave his phone near a GPU of Wohpe, believed to be fully contained. The AI discovers that could use the GPU itself as a resonator to establish an RF link with the phone, breaks the protocol, finds vulnerabilities and escapes into the Internet.

Habemus logo. From a collaboration between me (hand drawn pen and paper), AI (turn this shit into a logo) and Ben Gnomino (the human touch).

I just pushed a new version of DwarfStar that is likely to break stuff, too many new things: 1. micro-batching. 2. CUDA multi device support. 3. GLM5.2 4. Metal tensor parallel execution (via RDMA). 4. DSpark speculative execution. 5. Many other things probably.

It is always a good moment to remember that the recent results disprove another conjecture as well: the one of the many AI scientists that were LLM skeptics. People make errors, but remember that AI is not fully understood, and experts of the field make a lot of blatant errors.

I see very worrying US companies executives declarations depicting open weight models as a risk. How much sold to the most furious capitalistic vision you need to be, to really believe it is better that a few companies control AI for all the world? They said the same for OSS.