GitHub Repo
Ollama Community Reports Missing Text in LFM2 Output; Certain Python Strings May Disappear Silently
A community reproduction on Ollama 0.35.1 suggests that `lfm2:24b` may omit certain Python strings, altering JSON values and document content. Upstream has yet to confirm the issue, and there is not enough evidence to determine whether it stems from tokenizer decoding, model packaging, or the hardware path.
On October 4, the Ollama community reported an output integrity issue involving lfm2:24b: lowercase python may disappear in certain positions, while the form with a leading space displays normally. The report used Windows 11, Ollama 0.35.1, and a Radeon RX 9070 XT, with the model running entirely on the GPU. This does not establish that other platforms are affected. Issue report
The minimal reproduction uses /api/generate to prompt the model to continue from repeated JSON fragments, with zero temperature, non-streaming output, and raw:true. The reporter expected a language field containing python, but the actual response began at the following quotation mark, leaving an empty string. The official documentation says raw mode skips the prompt template. This makes the case a useful lead for investigating processing paths outside the template, but it does not directly prove which layer dropped the text. API documentation
The official model library lists model identifier d6c816d74887, matching the report, packaged as a Q4_K_M quantized version of about 14 GB. This lets follow-up tests target the same artifact rather than merely comparing models with the same name. The library also describes it as a hybrid architecture model suited to local deployment. Model details
One technical concern is that missing text can still leave behind parseable JSON. Based on this case, an application that checks only the format could accept a result whose content has changed, affecting language labels, code descriptions, or document indexing. Engineers can compare fixed strings and field values, and test variations in capitalization, leading spaces, and adjacent punctuation. These are validation approaches, not proven general fixes.
As of the time of review, the issue remains open, with no linked fix listed. The reporter suspects vocabulary mapping or the decoding process, but the page provides insufficient evidence to confirm the internal tokens involved. Next steps include checking whether upstream can reproduce the issue on CPU and other GPUs, comparing different runtimes with the same weights, and using regression tests to verify text integrity after a fix. Issue status