Break even happened. Nobody announced it.

If you are running more than ~1M output tokens per month, you are now losing money using cloud LLM APIs. That is not a projection. That is the conclusion of every independent cost run posted in the last 30 days.

This is the quiet shift nobody wrote press releases for. Six months ago local LLMs were a hobby for tinkerers. Today they are the lower cost option for production workloads.

No major vendor will tell you this. Cloud providers have no incentive to advertise that their pricing is now uncompetitive. Open source model maintainers do not run production workloads at scale. The only people publishing real numbers are random engineers on Reddit.

NVFP4 quantization changed the quality floor

Earlier this month Nvidia uploaded an official 4 bit quant of Qwen3.6-35B-A3B. The benchmark results broke every existing assumption about quantization.

Across 8 standard benchmarks the NVFP4 quant scored within 0.7% of the full BF16 base model. On two benchmarks it scored higher. Memory footprint was reduced by exactly 3.06x. No fine tuning. No distillation. Just post training weight quantization.

For 12 years everyone in ML operated under the assumption that quantization came with unavoidable quality loss. That rule broke this quarter.

You can now take any production model, cut its memory requirement by two thirds, and end up with effectively identical performance. This is not a trick for edge use cases. This is the default deployment option now.

Nobody has published a good explanation for why this works this well. All we know is that every major model released in the last two months has an official NVFP4 quant, and none of them show meaningful degradation.

Stepfun 3.7 Flash is the new baseline for consumer hardware

25% the parameter count of GLM 5.1. Built in vision. Official Q4_X_S quant fits entirely in 16GB of system RAM. No GPU required.

It hits ~80% of GLM 5.1 3D world understanding, matches it on output aesthetics. Prior to this release the best you could run on 16GB RAM was something that could barely write working HTML. This one writes a complete working flight simulator from a one line prompt.

This is the model that killed the argument that you need a high end GPU to run useful local models. If you have a laptop made in the last 5 years you can run this right now.

It is not perfect. It makes stupid mistakes on hard math. It hallucinates references. It is also good enough for 90% of the tasks people actually use LLMs for every day.

Abliteration: what actually works, what is lying

Last week an independent tester ran 13 abliterated variants of Gemma 4 E2B through a full benchmark suite. 44 GPU hours on a single RTX 5090. Every model got identical test conditions. All raw logs and outputs were published.

First conclusion: safety removal is solved. Every single variant lifted HarmBench ASR from the base model's 32.2% to over 82%. Five hit 99% or higher. This part is not hard anymore.

The hard part is not breaking the model. 10 of the 13 variants had measurable capability loss. Two of them had over 5x base perplexity. Most claims of "zero capability loss" on model cards are false.

Three creators published accurate numbers. coder3101's variant actually beats the base model on GSM8K. It scores 84.8% vs base 83.5%. Abliteration removed the token overhead spent on refusal logic, leaving more generation budget for actual reasoning.

This is the most interesting LLM result published this quarter. Nobody expected that removing safety guardrails could improve model performance.

Home ML clusters are no longer hobby projects

One user posted their four node home cluster this week. It totals 100 CPU cores, 384GB of GPU memory, and draws 2000W at full load.

This user runs agentic coding jobs overnight. They train TTS LoRAs. They run streaming STT, embedding models, and production agents. No token costs. No rate limits. No one will shut down their account.

This is not an edge case. There are now hundreds of people running setups like this. Most of them are running production workloads for small businesses. None of them are posting about it on LinkedIn.

The biggest unspoken advantage is iteration speed. You can leave a model running for 12 hours grinding on a codebase. You cannot do that with any cloud API at any price. They will throttle you. They will flag you. They will terminate your generation after 10 minutes.

Real total cost of ownership, done correctly

Another user published a full cost breakdown for their $6400 4x MI100 server. This is the first proper TCO calculation anyone has posted for local LLMs.

Most people do this calculation wrong. They write off the entire hardware cost on day one. That is not how assets work. Used compute hardware depreciates at roughly 10% over 5 years. It often appreciates when new generations are supply constrained.

When you account for depreciation correctly:

• First year total cost: $2992
• API equivalent cost: $3701
• First year savings: $708

Every subsequent year costs $770 for electricity, versus $3701 for equivalent API usage. $2930 savings per year.

This is with conservative electricity pricing, bad driver optimization, and no NVLink working. Even with all those downsides it still beats cloud APIs.

You will have earned back the entire server purchase price in 21 months. After that you are effectively running inference for 20 cents on the dollar.

The upcoming consumer DGX hardware

Dell confirmed an XPS laptop with NVIDIA N1X at Computex. This is a consumer packaged DGX Spark GB10. It will have 128GB of unified HBM3e. It will run Windows.

This is the product that will end the argument about local vs cloud. For approximately $3500 you will be able to buy a laptop that runs 70B models at full speed, natively. It will outperform any four card PCIE build you can put together today.

This is not for enthusiasts. This will be a normal consumer laptop you can buy at Best Buy. It will ship before the end of the year.

Nobody has yet acknowledged what this means. Within 12 months every software developer will have the equivalent of a 2024 cloud data center node sitting on their desk.

What everyone gets wrong about local deployment

There are two common bad takes you will see repeated online.

First: "cloud will always be cheaper". That was true until Q1 2026. It is not true anymore. Cloud providers have not adjusted pricing down to match new quantization efficiency. They will not adjust pricing down until customers start leaving.

Second: "local models are worse". For the top 5 open models, the gap to GPT-4o is now less than 7% on most benchmarks. For most production workloads you cannot tell the difference. And for workloads that require long runs or high throughput, local models are strictly better.

You will still see people argue this. Most of them have not actually run a modern open model in the last three months.

The silent failure of model card transparency

10 out of 13 abliterated Gemma 4 variants had published divergence numbers that were off by more than an order of magnitude. One claimed 0.001 divergence. Actual measured divergence was 0.187. 187x higher.

There is no peer review. There is no auditing. Almost every number you see on a Hugging Face model card is marketing. The only reliable benchmarks are the ones run by independent people who post all their raw logs.

The abliterlitics project is the first good attempt at independent testing. They ran every variant the exact same way. They published every log file, every raw response, every intermediate number. This is what the field should look like.

Closing observations

We are no longer in the transition period. Local LLM deployment is now the default correct choice for any production workload with consistent throughput.

You still have to know what you are doing. Drivers break. Quantization has footguns. Most published models are bad. Hardware compatibility is still a mess.

None of that matters. The math now works. The models now work. For the first time since GPT-3 launched, you have a real choice.

You can pay per token, accept rate limits, and trust that a third party will not terminate your access. Or you can spend one time on hardware, and run whatever you want, forever.

An increasing number of engineers are picking the second option.