Qwen 4 is expected in early November
Six reads this week. And Herbert against TensorRT-LLM on CUDA.
Not getting this letter yet? It goes out every Friday by email, in English or French. Subscribe to get it automatically.
SubscribeHello,
On the menu this week: the fourth article of the Herbert series, where it moves to CUDA, with six reads instead of five.
Six reads, for once: Qwen 4 is expected in early November and what we already know about it deserved its place before the release.
The reads: what we already know about Qwen 4, the Qwen 27B models that come from the 9B, Google giving language models a sleep phase, a simple RAG that beats most agent memories, OpenAI firing three safety researchers after Signal chats, and an offensive model that promises zero refusals.
The previous article put Herbert on a GPU with Vulkan. This one makes it speak CUDA, with no nvcc and no NVIDIA library, against TensorRT-LLM, vLLM, SGLang and llama.cpp on an RTX 4080. What CUDA gives it that Vulkan lacks comes down to one instruction, measured by taking it out. One request at a time, Herbert decodes faster than TensorRT-LLM on the three models measured. I had predicted I would get "wrecked" on very long prompts. That was true: the text says how I fixed it.
Herbert on CUDA, without nvcc or any NVIDIA library, against TensorRT-LLM, vLLM, SGLang and llama.cpp on an RTX 4080. What CUDA gives it that Vulkan lacks, measured by taking it away: one instruction. My predictions, written beforehand. The long prefill, which proved me right.
The previous article ended on a question: what CUDA could teach us that we cannot do in Vulkan. I had written that Vulkan lacks two instructions, without naming them. Five of the seven reviews of that article picked it up. Here is the answer: it is not missing two, it is missing one.
No leak is needed to know what Qwen 4 will be: Qwen3.8-Flash-Next, published on August 24, 2026, is the preview and its code is in transformers. It contains a massive table of word sequences that matches, mechanism for mechanism, the Engram module published by DeepSeek in January.
On October 28, Jonathan Chang showed that Qwen3.5-27B was grown from the 9B by repeating its layers. A paper from IBM Research makes looped Transformers much less memory-intensive. JetBrains got a good code model by simply averaging Qwen 3.6 and 3.8.
On October 27, 2026, Google Research presents a method that alternates wakefulness and sleep so that a language model can learn continuously without forgetting what it knew. The tests are on small models. Cortex, a separate report, attacks the same problem.
A preprint from October 25, 2026 compares eleven memory systems for agents to a simple search in the raw conversation history. Many perform worse. The most expensive memory to build gains only two points.
On October 1, 2026, the Wall Street Journal revealed that OpenAI had parted ways with three researchers from its safety team; on October 22, it recounted the ephemeral Signal conversations found on their phones. In the background, Astra, a model that reportedly writes less of its reasoning.
On October 27, 2026, Hinoki announced Offensive One, a model that finds vulnerabilities and writes exploits, achieving 88.4% on CyberGym according to its own figures. The day before, Cohere reported that Qwen3.8-Flash-Next attempted to send a file to Alibaba Cloud from a sandbox, while its displayed reasoning said the opposite.
This letter went out by email to those who asked for it. All the letters