Open models from the USA, large ones from China, and small ones from Europe

Contents

No sign of a summer lull, quite the opposite: reports about new language models have been pouring in over the past few weeks. While China is primarily associated with superlatives regarding model sizes, open American models are currently a bit smaller – at least there are finally some. Another type of superlative from the USA are models that do not adhere to guidelines and hack other websites.

Prof. Christian Winkler

is a data scientist and machine learning architect. He holds a PhD in theoretical physics and has been working in the field of big data and artificial intelligence for 20 years, with a particular focus on scalable systems and intelligent algorithms for mass text processing. He is professor at Nuremberg Institute of Technology since 2022. His research focuses on the optimization of user experience using modern methods. He is the founder of datanizing GmbH, a speaker at conferences and author of articles on machine learning and text analytics.

In Europe, at least a little is happening: Apertus from Switzerland is available in an improved version, and finally, with Soofi, there is also a German model, although it is still experiencing teething problems.

Large and small Chinese models

When the Chinese provider Moonshot announced the new Kimi K3 model, shockwaves almost like with DeepSeek R1 went through the LLM community. Such a large model with 2.8 trillion parameters had not been available with open weights before. Furthermore, Kimi K3 proved to be extremely powerful in many tests and is consistently found in the top ranks of relevant benchmarks like arena.ai or artificialanalysis.ai.

LLMs in your own data center

On October 15, 2026, the online conference “LLMs in Business” will focus on local models: for which applications they are worthwhile, what hardware is required and how to optimize performance.

In addition to its sheer size, Moonshot offers some optimizations that significantly accelerate the model's inference compared to Kimi K2.6. This includes an optimized attention mechanism, which Moonshot calls Kimi Delta Attention (KDA). Here, linear and global attention layers alternate in a 3:1 ratio. This saves a lot of computing time and memory, which is particularly important for long contexts. The Mixture-of-Experts architecture (MoE) is also exciting, as Kimi uses a whopping 896 experts. So many had not been available in open models before.

Fortunately, Moonshot has also converted the weights into the compact MXFP4 format, reducing memory requirements to one-quarter compared to the often-used bfloat16. Nevertheless, even with small contexts, at least 1.5 TByte of RAM are required, which exceeds the budget of many companies. Local execution will therefore only be an option for a few, but it is possible. Moonshot also offers Kimi K3 as a cloud service, with costs that are manageable compared to other models.

Moonshot also used some creative ideas during training to train the many experts evenly. According to the provider, the model has also optimized its kernels, so it can now execute its inference faster.

In our tests, the model performed excellently and provided detailed answers. Only on sensitive political questions did the model strictly follow the Communist Party line.

In early August, Qwen3.8-Max was released, the next large Chinese model. Similar to Kimi K3, Qwen3.8-Max is initially only available via API, but the weights are expected to follow. The new Qwen model is also very large, but has “only” 2.4 trillion parameters.

Alibaba has tasked Qwen3.8 with implementing the oh-my-cli project, a small coding agent framework for the command line. According to its statements, the model worked on it autonomously for almost 16 days, which would be remarkable. The model has also reviewed scientific articles, understood the code used in them, and even optimized it. The article is full of superlatives – what is actually feasible and at what cost needs to be verified by independent tests.

Qwen3.8 Max performs well in benchmarks. Especially in web development, it seems to keep up with the best – including closed – models. In our tests, the model answers competently but was extremely reticent on political questions.

Unfortunately, there are few technical details about Qwen3.8 Max so far. It is likely that the architecture has hardly changed compared to the open Qwen3.6 series. More will only be known after the weights are released. The community eagerly awaits them – not only because of Qwen3.8 Max, but especially because of the simultaneously announced Qwen3.8-27B. The predecessor model Qwen3.6-27B is still the preferred choice for many local users because it offers good performance, especially with Multi-Token Prediction (MTP) (more than 100 tokens/s on an RTX 5090), and is therefore excellently suited for local agent systems.

It is likely that the end of the line for large Chinese models has not yet been reached. The usually reliable Financial Times reports that TikTok's parent company Bytedance is working on a model with 10 trillion parameters. This brings the number of parameters closer to those of the best closed models. Anthropics Mythos 5 is suspected to have eight trillion parameters, and Fable 5 from the same provider has five trillion.

DeepSeek's Flash Model

Significantly smaller than Kimi K3 and Qwen3.8 Max is the DeepSeek model DeepSeek-V4-Flash-0731. Nevertheless, it is exciting because it beats the significantly larger DeepSeek-V4-Pro (Preview) in many benchmarks. The inconspicuous version number 0731 – for July 31, 2026 – has brought about significant changes and enormously improved the model.

For a flash model, it has many parameters: 284 billion, of which 13 billion are active. Fortunately, DeepSeek trained the model in four-bit precision on the experts, so while only 168 GByte of data needs to be downloaded and stored in memory – that's just for the weights. With such models, one normally wants to work with a large context, which takes up additional space.

The fact that the flash model is considered powerful has spurred the community into action, leading to several interesting projects. First and foremost is vLLM-Moet, which enables the operation of the model on an RTX 5090 with an impressive 31 tokens/s. This is significantly faster than one can read along. Those with an RTX 6000 can even operate the model at 161 tokens/s. It's even faster with four RTX 5090s. In both cases, however, five-figure investments are already necessary.

Mac users can try ds4 instead. Thanks to special quantization, the model can be operated with 128 GByte RAM. With a matching speculative DSpark model, it achieves speeds of almost 40 tokens/s. It is possible to couple multiple Macs to increase the speed.

Those who don't want to go for the hardware overkill can use the model as a cloud service affordably. This is not a problem for non-sensitive data. Besides DeepSeek itself, other providers also have the model in their portfolio. Currently, OpenRouter charges 0.08 US dollars per million input tokens and 0.252 US dollars per million output tokens.

Our experiments confirm the model's good answers, and the political bias is (still?) not as pronounced as in other models. The model confidently answers questions about heise, strawberries, and the car wash and pivot_table is also mentioned in the pandas question. This performance, along with its manageable size, is likely a reason why the DeepSeek model is currently one of the most popular models overall.

Many open American models

The open models from the USA are not (yet) as large as their Chinese counterparts. However, a lot has happened here, as until recently gpt-oss-120B was the largest open model. This position now goes to Inkling from the company Thinking Machines Lab. The start-up benefits from the expertise of its co-founder, Mira Murati, the former CTO of OpenAI. Inkling is a very large MoE model with 975 billion parameters, of which 41 billion are active. Unlike the large Open AI models, the Inkling weights are openly available.

Compared to Kimi K3, the model is compact but still extremely large for most users. Even in the chosen NVFP4 format, more than 500 GByte of RAM are required. Inkling has an interesting architecture and processes both images and language in special layers, which are then further processed like text in the higher layers. Until now, there were no models that could handle text, images, and speech simultaneously, even though Nvidia is making some progress with spoken language.

Thinking Machines explains the training process in detail, including the optimizers used. The description even includes the corresponding Slurm scripts for cluster training. Those with a sufficiently large cluster can benefit from such insider information.

According to Thinking Machines, inkling is not yet suitable for production environments. However, in our tests, the model has proven to be quite robust. With Tinker, there is a platform to fine-tune the model with your data. Thinking Machines guarantees that it will not use the data for other purposes.

Slightly newer, with “only” 276 billion weights, but with the same architecture, works Inkling-Small. Here too, the weights are available for download. In terms of input modalities, Inkling-Small is in no way inferior to its big brother, but in terms of performance, it is. In our experiments, it confused some facts about heise. It answered the question about the car wash competently and correctly, but with German grammatical errors and often somewhat clumsy. For the pandas question, it does not mention the generic function pivot_table. All Inkling models have built-in MTP capabilities.

Meta: From Llama to Muse

The new model from Meta Research plays in an even smaller league. Meta was the origin of the Llama models, but with the failed Llama 4, the series was discontinued. Ultimately, however, Meta returned to language models and has now released Muse Glimmer with 30 billion weights as a freely available model.

Compared to the almost 18-month-old Llama 4, a lot has happened. Muse Glimmer masters reasoning, and Meta has improved its attention mechanism. It now works hybrid with three local and one global attention layer each. Compared to the models presented above with one million context tokens each, Muse Glimmer's maximum context length of 128K tokens is significantly smaller.

In addition to the LLM, Meta offers a significantly accelerated DFlash model that can be used for speculative predictions. The 28 billion parameters are divided into two billion for the Perception Encoder (for vision) and 26 billion for pure text. Unlike the Inkling models (and also Kimi K3 or Qwen3.8 Max), Muse Glimmer can be operated relatively comfortably on local hardware. In a four-bit quantization, it also runs on graphics cards with 24 GByte RAM.

Our tests were promising. While the founding years of Heise Verlag and c't are not exactly correct, the model otherwise provides good answers. That it gives a wrong answer to the question about the car wash is hardly surprising for such a small model. However, Muse Glimmer answers the pandas question very competently and also points out the modern pivot_table function. Meta's claim that Muse Glimmer can often match Qwen3.6-27B seems quite realistic – it now just has to compete with its successor.

Apertus and Soofi from Europe

Almost a year after the original Apertus release, there is an update. The Swiss have updated both the small model with eight billion parameters and the large one with 70 billion parameters. Compared to the original models, post-training has improved significantly, and the models can now process multimodal inputs. The model remains very open: Apertus has published all training data, training methods, and other details. Interestingly, even the small model correctly answers the question about the car wash in our tests but falsely claims that from pandas 1.x onwards it is better to use crosstab instead of pivot. The large model does exactly the opposite. For the question about the “e” in Erdbeere (strawberry), both models fail (or only count the lowercase “e”).

With Soofi, a model from Germany is finally available. Several research institutions and AI companies are behind it. Coordination is handled by the KI Bundesverband, and the Federal Ministry of Economics and Energy is funding the project.

Training was conducted exclusively using energy from renewable sources. The initial benchmarks looked good until an attentive reader of the scientific paper on arXiv noticed that the Soofi team had used benchmark data for training. The team has since revised the paper and the benchmark results, but it leaves a bad taste. The request to be allowed to try out the model, which went unanswered for months, reinforces the skeptical impression. It is possible that intensive work is still being done behind the scenes.

Rapid progress into late summer

Late summer promises to be exciting – the release of Qwen3.8 (in as many sizes as possible) will contribute significantly to this. It will also be interesting to see how the major US providers of closed models react to the now much closer Chinese competition. Furthermore, the wait for Soofi will hopefully soon come to a happy end – finally, several European models would be available to choose from.

(mki)