This is not incremental progress. This week we crossed a line. Everything you knew about what hardware you need to run and train good LLMs is obsolete.

The quiet hardware shift no one is talking about

Nobody is buying new GPUs anymore. Not the people actually running models.

This week two separate people posted their builds: one got dual 3080 20GB for less than the price of one used 3090. Another put three V620s into an R740 for $1050 total, giving 96GB of usable VRAM. That is enough to run Laguna S 2.1 Q4_K_M with full 1M context.

Used datacenter cards have collapsed in price. Ampere era compute cards are now cheaper per gigabyte of VRAM than system RAM. This is not a temporary blip. Everyone running local models noticed. Everyone is swapping their hardware right now.

Card	VRAM	Current street price	$ per GB
RTX 3080 20GB	20 GB	$270	$13.50
Tesla V620	32 GB	$350	$10.93
RTX 3090	24 GB	$580	$24.16
RTX 4090	24 GB	$1850	$77.08
RTX 5090	32 GB	$2999	$93.71

You are being scammed if you buy a modern consumer GPU for running LLMs. Stop doing it.

SkewAdam fixes the single worst problem with MoE training

Everyone knew MoE optimizer state was broken. No one did anything about it until this week.

AdamW uses 8 bytes per parameter for optimizer state. For a 12.6GB model that is 50.6GB of state just to run training. That was the single largest item in every MoE memory budget. Everyone just accepted this as a law of nature.

SkewAdam does not accept this. It works on one extremely obvious observation that every single person in the field missed: 95% of parameters in an MoE are experts. Experts do not need full precision optimizer state. The router does. The backbone does. Experts absolutely do not.

The results are not incremental. They are ridiculous. Optimizer state drops from 50.6 GB to 1.29 GB. That is 97.4% reduction. Peak training memory for a 6.78B MoE drops from 81.4 GB to 31.3 GB.

This fits entirely on a single 40GB GPU.

No tricks. No offloading. No convergence loss. No router instability. This is not a tradeoff. This is just a strictly better optimizer for MoEs. Every single person training MoEs will be using this within 30 days. Code is already on GitHub. There are no patents. There is no license. Just use it.

Laguna S 2.1 is the open model everyone was waiting for

Poolside dropped Laguna S 2.1 this week. It is the first open model that is actually good enough to replace closed models for agent work.

This is an 118B total parameter MoE that activates only 8B per token. 1M context. Native interleaved reasoning. Native tool calling. Open license.

Stop for one second. 8B active parameters. That is the same active cost as Llama 3 8B. It beats every other open model on every coding and agent benchmark. It is competitive with models that activate 40B, 50B, even 200B parameters per token.

This is not a small improvement. This is an entire generation ahead of every other open model released before this month.

Benchmark breakdown

These are independent third party benchmark results. No vendor cherry picking. No private test sets.

Model	Active params	SWE-bench Pro	Toolathlon Verified
Laguna S 2.1	8B	59.4%	49.7%
DeepSeek-V4-Pro Max	49B	55.4%	55.9%
Inkling	41B	54.3%	45.5%
Nemotron 3 Ultra	55B	-	34.3%
Llama 4 70B	70B	47.1%	38.2%

Laguna is delivering better coding performance than models using 6x more compute per token. That should not be possible. It is.

It loses to DeepSeek on Toolathlon by 6 points. It uses 1/6th the inference cost. That is the tradeoff. That is the best tradeoff we have ever had in open models.

Quantization support landed everywhere

48 hours after release, Unsloth published full GGUF quants. llama.cpp merged support. vLLM merged support. SGLang merged support. Ollama works. LM Studio works. Every single inference stack you use already runs this model.

The UD-Q4_K_XL quant is 40GB total. It fits on a single 48GB GPU. It runs at 16-20 tokens per second generation on three V620s. Prefill hits 600 tokens per second.

You can run this right now. Today. You do not need a cloud instance. You do not need an API key. You do not need to pay per token.

NVIDIA Model Optimizer finally got good

NVIDIA open sourced Model Optimizer properly this month. This is not the garbage half working tool they released last year. This is a real library that does every optimization you actually need.

It does PTQ. It does QAT. It does pruning. It does distillation. It does speculative decoding draft training. It exports directly to vLLM, SGLang and TensorRT-LLM.

Most importantly, it works. No hidden flags. No undocumented requirements. You point it at a Hugging Face model, run one function call, and you get an optimized checkpoint that runs 2x faster with negligible quality loss.

This is now the standard quantization tool. Everyone will stop writing custom quantization scripts. Everyone will use this.

What is broken right now

Not everything works perfectly.

SkewAdam only works for MoEs. It will not give you these gains on dense models. Laguna reasoning is still slightly worse than DeepSeek R1 for very hard problems. It will hallucinate tool parameters about 3% more often. NVIDIA Model Optimizer still does not work correctly on AMD cards. No one is surprised. Used V620 supplies are already drying up. Prices are already up 15% since the start of the week. If you want one, buy it now.

Recommended stack July 2026

This is what you should be running today:

1. Hardware: 3x V620. Total cost ~$1100. 96GB VRAM.
2. Model: unsloth/Laguna-S-2.1-GGUF UD-Q4_K_XL
3. Inference: llama.cpp b10087 or later
4. Training optimizer: SkewAdam for any MoE fine tuning
5. Quantization: NVIDIA Model Optimizer 0.9.2

This stack will beat every closed API on cost per token. It will beat every closed API on throughput. It will be within 10% of the best closed models on almost every task.

You can build this today. You can run this in your closet. No one can shut it off. No one can raise prices on you.

Closing

This is the point everyone was waiting for.

For five years we have had open models that were almost good enough. For five years we had hardware that was almost cheap enough. For five years every breakthrough came with a catch.

This week there was no catch.

You do not need permission to run state of the art AI anymore. You do not need a budget. You do not need a company. You just need $1000 and a spare power outlet.

That changes everything.