The one month test

Day one release threads are noise.

Forty people run their favourite prompt. Half declare the new model the best thing ever built. Half call it garbage. None of this survives two weeks of actual work.

No one posts the follow up. No one writes the thread three weeks later saying they quietly uninstalled it. No one admits they went back to the old model they swore they were leaving.

This is the most important filter in the entire open LLM space. A model you stop thinking about is the highest praise available.

What actually survived production this quarter

Over the last 6 weeks this is what people are still running, aggregated across hundreds of deployment reports and comment threads:

Model	Primary use case	Still installed after 30 days
Qwen 3.6 27B	General purpose	82%
Qwen 3 Coder Next	Code generation & debugging	79%
Ling 3.0 Flash	Agent tool execution	61%
Gemma 4 31B	Multimodal baseline	47%
GLM 5.2	Long context workloads	38%

Every other major release from the last quarter is already gone. Most were uninstalled within 10 days.

No 70B class model made this list. No 120B model made this list. All of the top entries are under 30B parameters.

The quiet dominance of Qwen under 120B

No one planned this. No marketing team pushed this. It just happened.

Engineers keep testing every new release. They run benchmarks. They do side by side comparisons. Then they go back to Qwen.

This is not fanboyism. This is exhaustion. People have work to do. They will stop testing the second they find something that does not break, does not hallucinate on trivial tasks, does not surprise them with stupid failures at 2AM.

Right now for anything under 120B parameters, there is no credible general purpose alternative. There is no credible coding alternative. Every release for the last 5 months has failed to displace it.

This is not a permanent state. It is the state right now.

PerceptionBench: The first good multimodal benchmark

For two years every multimodal benchmark was useless. All of them measured how well models had memorized the test set. None measured actual visual perception.

That changed this month with PerceptionBench.

The authors did something very simple. They took every failure from every existing benchmark, broke them down into atomic failure modes, then built 3000 test questions that isolate exactly one perceptual capability each. No reasoning. No trick questions. Just can the model actually see what is in the image.

No model scored over 60%.

Rank	Model	Overall Accuracy	Hallucination Resistance	OCR Accuracy
1	GPT-5.6-Sol	59.7	26.9	54.9
2	Kimi K3	58.5	41.7	61.2
3	Claude-Fable-5	57.2	45.0	64.3
6	Seed 2.1 Pro	55.0	49.8	66.7
13	Gemma 4 31B	40.7	32.1	46.7
16	GLM 4.6V	32.5	26.9	39.2

The most important finding: every model has catastrophic blind spots. Two models can have identical overall scores while being completely useless at entirely different things.

This benchmark will replace every other multimodal evaluation for production teams. It already has for most people running private deployments.

Corporate releases: Trust is breaking

Microsoft pulled Mage-Flow 12 hours after release. The Hugging Face pages return 404. No explanation. No announcement.

This is no longer an anomaly. This is standard operating procedure.

Google has pulled three models in the last six months. Anthropic has restricted access to two open weights releases after publication. Elon Musk promised Grok 3 would be open source one year ago. It never happened.

Teams can not build production systems on this. You can not schedule engineering work, run validation, write integration code, for a model that might get deleted tomorrow.

This is the single largest opportunity for the open community right now. Corporate actors have completely abandoned the trust that made people use their releases. No one will give them the benefit of the doubt any more.

The new community pipeline

Almost all useful work now happens outside official corporate releases.

Notice the corporate entity is only involved at the very first step. All validation, all hardening, all useful adaptation is done by random people on Hugging Face with no affiliation.

This is not a bug. This is now the actual production pipeline.

The silent work no one posts about

No one makes front page posts for these things. But this is what actually moves the ecosystem forward right now:

• ianncity published 50,000 clean reasoning traces distilled from GLM 5.2. No announcement. Just uploaded. This dataset is already being used to fine tune half a dozen smaller models.
• unstonio published 218,000 tagged 24x24 pixel art samples with full taxonomy and quality scores. It is the only clean pixel art dataset in existence.
• owensong released Inflect Micro v2, a 1B parameter model that does agent tool calling better than most 70B models. It has 12 likes on Hugging Face.
• baidu republished Unlimited OCR, fixed the broken dependencies, and did not mention it anywhere.

None of these got day one hype threads. All of them are already in use in thousands of deployments.

What this means for teams evaluating models

If you are selecting models for production right now:

Ignore every launch day thread. Ignore every official benchmark. Ignore every press release.

Wait two weeks. Then check if anyone is still running it. Wait another two weeks. Then check again.

If after 30 days people are still using it, and no one is complaining about silent failures, you can start testing it.

This is the only filter that works. Everything else is marketing.

The state of the open ecosystem right now

The open LLM ecosystem has matured. It is no longer about who can release the biggest model first. It is about who can release something that people will still be running next month.

Right now the community is winning on reliability. It is winning on data. It is winning on trust.

Corporate actors still have the biggest models. They still have the best raw benchmark scores. But they have completely abandoned the one thing that actually matters for adoption: consistency.

Teams will choose a slightly worse model that is always there over a slightly better model that might get deleted tomorrow. They already are.

That is the quiet shift no one is writing press releases about. That is the state of the open LLM world right now.