Reads an entire book in one prompt and answers questions about it

You put the whole document into the prompt and ask. Nothing gets chopped into chunks or looked up from a search index; the model reads all of it, up to 512K tokens, which is roughly 350,000 words, a long novel with room to spare. It thinks before it answers, with a per-request dial for how hard: low, medium or high. High is what the makers recommend and it can run long, so a detailed question over a big document takes a while but comes back with the reasoning visible.

It is a general model too, not just a long reader. It handles tool calls for agent work, and IFM, the group behind the LLM360 open-training project, puts it ahead of Gemma 4 12B and Qwen3.5 9B on their long-context reasoning, competition math and software-engineering tests. Those are the makers' own numbers; the technical report is due at the end of September.

The unusual part is how open the release is. Pretraining data, training logs and around fifty intermediate checkpoints are public, from the first pretraining stage through the context extension and the final tuning, so you can load any point in training and watch a capability appear.

There is no official demo, and the community Space for the 7B was paused when we checked, so the realistic path is running it yourself. The weights are Apache 2.0; IFM ships a single 18 GB GGUF file for llama.cpp-style tools, plus one-line vLLM and SGLang recipes. The weights fit a 24 GB card, but the context you can actually use depends on memory: the full 512K window needs data-center hardware, so start with something like a 100-page report. First test: convert a long PDF to text, paste it in, ask about a detail from the middle, and check whether the answer quotes the right passage. Set reasoning to high and allow a long reply.

Long context is memory hungry: half a million tokens on your own hardware means a large GPU, and just reading that much text takes minutes. With no demo running today, you need a machine to try it at all. At high reasoning effort the thinking can run very long, so budget output tokens or it gets cut off mid-answer.

• Ask questions about an entire contract, thesis or manual without splitting it up first.
• Drop a small codebase in and ask where a specific behaviour is implemented.
• Summarize a year of support tickets or meeting notes in one pass.
• Load an intermediate checkpoint and compare how it handles the same prompt before and after context extension.

Try the demo

View model page

Read this issue on neuralexpedition.com →