You are wasting your time testing 30B class models released in the last 30 days. Almost every one of them is broken. That is the single most important takeaway from this week in the open LLM community.

Everything else this week flows from that observation: people are quietly abandoning the 27-33B size bracket, moving back up to 120B MoEs, building workarounds for broken chat templates, and inventing new inference features to compensate for model defects.

Samsung just changed every inference cost calculation you have

This is not an LLM release but it is the single most important thing that happened this week. Samsung's chip division will post more profit in 2026 than it made in the entire prior 40 years combined. That is not a typo. Quarterly profit increased 19x. They passed Nvidia this quarter as the most profitable company on earth.

Every single person running local models has felt this already. H100 prices are up 42% since March. DDR5 ECC is up 61%. NVLink modules are backordered 14 weeks. This is not temporary. Memory manufacturers have coordinated supply perfectly. For the next 18 months, every inference optimization, every quantization, every model size choice matters 2x more than it did 6 months ago.

You will not get cheaper hardware. You will get better models. That is the only path forward.

The 30B model dead zone

Nobody is saying this out loud but everyone has noticed. Every dense and MoE model released between 26B and 35B parameters in the last two months is defective.

Qwen 3.6 27B falls apart on multi step tool calls. It will generate spaghetti architecture, ignore separation of concerns, produce god classes that violate every single principle of software engineering even when explicitly instructed not to. It will happily write code that works once and can never be modified.

Gemma 4 26B and 31B cannot make decisions. They will ask for confirmation after every single step. You will spend more time typing "ok" than you would have done the task yourself.

All of them have broken chat templates. All of them regress on long context. All of them produce blank responses 7-12% of the time when doing tool calling.

This is not bad luck. This is the result of every model team cutting alignment training to hit reasoning benchmarks. They optimized exclusively for MMLU and GSM8K. They broke every other capability that matters for actual use.

120B MoE is the new working sweet spot

If you have 128GB of system RAM or 48GB of VRAM, stop testing anything else. Qwen 3.5 122B works. That is the consensus this week.

It does not get confused. It does not ask for constant confirmation. It will not generate garbage architecture. It will complete 8 hour unattended agent runs. It will extract 160 powerpoints correctly on the first try.

This is the first size bracket where teams did not completely break the model to beat benchmark leaderboards. There is a very clear capability cliff right now. Everything below 100B active parameters is optimized for benchmarks. Everything above is still optimized for actual use.

You do not need 8 GPUs. You can run this quantized at IQ4_XS on a single RTX 4090. It will run at ~18 tps. That is fast enough.

preserve_thinking is the most important new feature no one announced

This week three separate model releases added the same unadvertised feature. tarruda's Step 3.7 Flash quant, ThinkingCap Qwen 3.6, and Reasoning-Medical 27B all implement preserve_thinking.

Instead of discarding the model's internal reasoning trace after every reply, this flag retains the full <think> blocks in the conversation history.

This is not a minor quality of life change. This cuts redundant reasoning by 60% on long conversations. It makes agent behaviour consistent. It improves KV cache hit rate by 35%. It removes the single largest cause of model drift over 10+ turn conversations.

No model vendor announced this. No paper was published. Three independent people just built the same fix at the exact same time, because everyone was hitting exactly the same failure mode. This is how progress actually happens in the open LLM space.

Reasoning-Medical 27B is the first good domain fine tune

Most domain fine tunes are just base models with 1000 flashcard examples dumped into LoRA training. This one is different.

EpistemeAI released Reasoning-Medical 0.1 this week, trained on 100,000 curated medical reasoning examples with GRPO. It beats MedGemma by 11 points on MedQA. It hits 96.9% accuracy.

Benchmark	Reasoning-Medical 0.1 27B	MedGemma 1 27B
MedQA 2 shot	96.9%	85.3%
MMLU-Pro Biology	93.0%	78.1%
MMLU-ProX Biology	96.8%	74.2%

It also implements preserve_thinking correctly. It supports native multimodal input for radiology and pathology images. This is the first open medical model that clinicians are actually starting to use for draft note taking.

It is also the first model that properly implements gated access for safety without making you fill out a 20 page form. You ask for access. They grant it. That is it.

GLM 5.2 and the return of open source fearmongering

This week Futurism ran an article claiming GLM 5.2 is a dangerous cybersecurity tool that should be restricted. The entire open LLM community called this out immediately for what it is: pre-justification for export controls and model registration laws.

Semgrep ran their own benchmarks and titled the post "We Have Mythos At Home". GLM 5.2 is good. It is not magic. It will not let every script kiddie break every piece of software on the internet.

The important thing here is that this argument is now happening. For the last three years open models flew under the radar. That is over. Every good open model release from now on will be followed by a press cycle arguing it is too dangerous to allow.

MTPLX V2 hits 82 TPS on Apple Silicon

If you run models on MacOS you can stop using default MLX. MTPLX V2 was released this week and it is not even close.

82 tokens per second on Qwen 3.6 27B on an M5 Max. That is faster than most people get on an RTX 4090.

They did this by writing custom quantized matmul kernels specifically for verification passes, not just generation. They also rewrote the SSD KV cache implementation to eliminate 90% of the stalls that happen on 128k+ context runs.

This is the single largest inference speed improvement on consumer hardware in the last 12 months. No one at Apple did this. One independent developer did this.

Openness scores actually mean something now

Artificial Analysis published their updated openness index this week. For the first time this is not a meaningless list.

Model	Openness Score	Training data published	Training code published	Weights only
K2 Think v2	97	✅	✅	❌
Qwen 3.5	62	❌	❌	✅
DeepSeek V3	41	❌	❌	✅
Mistral Large 2	28	❌	❌	✅

Most people in the community still do not care about this. They should. Right now every model you use is a black box. You have no idea what is in the training data. You have no idea why it behaves the way it does. You have no way to fix it when it breaks.

K2 is the first frontier model that anyone can reproduce. That is a much bigger deal than 2 extra points on MMLU.

Closing observations

This was a quiet week. No big vendor announcements. No model that claimed to beat GPT-4o.

This was the week where everyone stopped paying attention to the leaderboards. This was the week where people stopped testing every new model that gets posted, and started building fixes for the ones that almost work.

That is good. That is how this field matures.

Benchmarks are dead. Actual use is all that matters now.