We are past the point where open source LLMs are toys. This week we crossed a quiet threshold: you can run a frontier class 295B parameter MoE model on a stock laptop. Almost no technical press covered it. Most of the important work this week happened not in press releases, but in forum posts, broken dataset viewers, and GGUF commit threads.

Tencent HY3 runs 295B parameters on 128GB laptop

This is the big one. Tencent dropped HY3 295B-A21B last week, a MoE with 295B total parameters and 21B active per token. One week later, a user had it running at 32 tokens/second on an M5 Max Macbook. That is not a typo.

Test condition	Tokens / sec
Prefill 512, empty context	528
Decode 128, empty context	32.4
Prefill 512, 16K context	124
Decode 128, 16K context	16.3
Decode with MTP enabled	38.1

This is faster than most people were running 70B dense models 12 months ago. The entire quant fits in 107GB. You do not need a server. You do not need a datacenter GPU. You need a laptop that you can buy today at an Apple store.

The installation process is still messy. You have to patch llama.cpp, adjust the GPU memory limit, and manually edit the architecture string in the GGUF header. None of that stopped people. As of writing there are 11 different public quants on Hugging Face.

The MoE value argument is finally being resolved

For 18 months the default take on LocalLLaMA has been: MoEs are scams. A 128B MoE with 12B active is just a bad 12B dense. This week that consensus broke.

The core mistake everyone was making is treating active parameters as the upper bound on model capability. That was true for first generation MoEs with terrible routers. It is not true anymore. Modern routers do not just pick random experts. They route tokens to the subset of parameters that have learned the exact pattern required for that token.

A 295B MoE with 21B active is not a 21B model. It is a model that gets to choose which 21B parameters it uses for every single token, out of a pool of 295B that have all seen different training data. That is a fundamentally different capability. You cannot replicate that with any dense model of any size.

This is not theoretical. HY3 is beating DeepSeek V4 Flash 671B on most benchmarks, while using one third the active parameters and running twice as fast.

Nobody is talking about the good dataset releases

Model releases get all the upvotes. Datasets are where all the actual progress happens. This week three extremely high quality datasets dropped with almost zero discussion:

Dataset	Size	License	Notes
kyutai/rocket-science	27 TB	CC BY-NC-SA 4.0	Perfectly aligned Rocket League gameplay, frame state, and player actions. This is the best world model training dataset ever released publicly.
LiquidAI/antidoom-mix-v1.0	~11M rows	Mixed permissive	Cleaned instruction tuning mix with full provenance per row, zero eval contamination. This fixes the single biggest problem with every public tuning dataset.
netflix/Vera-Layered-Video-Dataset	18k samples	Mixed	Layered video editing ground truth. Every sample has separate foreground, alpha, and background layers.

Antidoom is the most important one. Every popular instruction dataset right now has confirmed eval contamination. Every benchmark result published in the last 6 months is suspect. Antidoom explicitly tracks origin, split, and heldout policy for every single row. You can train on this dataset and trust your eval results. That alone makes it the most important release this month.

Local inference operating cost is now the dominant metric

Nobody talks about electricity cost. They should.

A user posted running cost numbers for an Asus Strix Halo this week. Full load 24/7, running inference, compiling models, and hosting services: $0.48 per day. That is $14.40 per month.

An A6000 draws 300W just for the GPU. That is $2.16 per day, $64.80 per month. And that is before CPU, RAM, cooling, or idle power.

For most production use cases you do not need 200 tokens per second. You need 10-15. You can get that on hardware that costs less to run than a single OpenAI API key. This is the argument that nobody at Nvidia will ever tell you. For 95% of deployments, throughput is no longer the bottleneck. Operating cost is.

The antidoom dataset design is correct

Look at the schema for antidoom. Every row has:

• Unique stable id
• Full original source
• Exact split from original source
• Original license
• Original sample id
• Explicit heldout policy

This is how every training dataset should be structured. Right now almost all datasets are just piles of prompts. You have no idea where they came from, what split they were on, or if they have already been leaked into every benchmark.

Antidoom does not just give you good data. It gives you audit trails. You can go back to the original source for every single prompt. You can verify that none of your training data appears in your test set. This is not a nice to have. This is table stakes for doing any kind of legitimate LLM research.

We are already nostalgic for Bloom

Someone posted a link to Bloomz-p3 this week. It got 800 upvotes. That model is almost four years old.

Bloom was the last big open LLM that everyone built together. It was slow. It was bad at most things. Nobody uses it anymore. But everyone remembers it.

Every model released now is built by one company. There are no more 1000 researcher collaborations. There are no more public training logs. There are no more open development processes. We got better models. We lost the thing that made this space fun.

Time capsule LLMs are the best kind of weird project

One user released a 500M parameter model trained exclusively on text from 1800-1875. No modern data. No alignment. Just 40B tokens of newspapers, books, letters, and government documents from 150 years ago.

It writes perfect Victorian prose. It can give you authentic 1850s plum pudding recipes. It has absolutely no idea what a computer is. It will never give you a corporate HR answer. It will never lecture you.

This is the best use case for open source LLMs. Nobody at OpenAI will ever train this model. Nobody at Google will ever release this model. This is the stuff you can only do when you can run the entire training pipeline on your own hardware.

Stop evaluating MoEs wrong

The entire conversation about MoE quality is using the wrong metric. People keep comparing total parameters. They keep comparing active parameters. Neither tells you anything useful.

The only number that matters for a MoE is router agreement. What percentage of the time would a human judge agree that the model picked the correct expert for that token?

Good MoEs have router agreement above 82%. Bad MoEs are below 65%. There is no other number that comes close to predicting real world performance. Until people start publishing this number, every argument about MoE vs dense is just noise.

Closing observations

This was a normal week. There was no big announcement. There was no viral demo. But this was the week that it became obvious that open source has already won.

You can run a frontier model on a laptop. You can train clean models on uncontaminated data. You can build things that no large company will ever build. All of this works right now. Most people still have not noticed.

The next twelve months will not be about making models bigger. They will be about making them cheaper, quieter, and more specific. The teams that win will not be the ones that can run the largest model. They will be the teams that understand that you no longer need the largest model at all.