Mistral Medium 3.5 128b ggufs are fixed
Broken GGUF files caused garbled outputs; now fixed for stable performance.
Mistral Medium 3.5, a 128-billion-parameter language model, had its GGUF quantized files broken, causing severe output degradation particularly at long context lengths. The issue, reported by users on Reddit, stemmed from a corruption in the quantization process that produced garbled, unstable text even at shorter contexts. Unsloth, the team behind the GGUF conversion, acknowledged the bug and released a fix.
According to the Reddit post from user Sunija_Dev, the updated GGUF files are "a lot more stable, even at short context." Previously, using the wrong prompt format would quickly devolve into gibberish, but the corrected version no longer exhibits this fragility. The fix is documented on Hugging Face, where Unsloth announced the resolution in a discussion thread. For professionals running Mistral Medium 3.5 locally via llama.cpp or similar tools, this update restores reliable performance across all context lengths.
- All GGUF quantized files of Mistral Medium 3.5 128B were producing bad outputs, especially at long context.
- Unsloth announced the fix on Hugging Face, addressing the prompt format sensitivity that caused gibberish.
- Updated version shows significantly improved stability even at short context, fixing the underlying quantization bug.
Why It Matters
Reliable local inference of 128B models depends on correct GGUF files; this fix restores production-grade stability.