March 3, 2026 updates ?
@danielhanchen - You updated almost all of the GGUFs again today 3/3/26 but left out Qwen3.5-35B-A3B-UD-Q5_K_XL.gguf. It this intentional or an oversight? Also, what's changed today compared to the 2/27/26 uploads? Qwen 3.5's core models are great, but it would probably be better to take an extra day and have a higher quality release. Just a pattern on GGUF quants in the last 3-4 months from unsloth.
Separately, the deep dives into KLD, perplexity, accuracy across various quantizers is great stuff. Hopefully we see more such analysis for the more notable model quants from now on
Q4_K_M is also disappeared
Hey sorry - we're also doing 122B, 395B as well - 35B now reuses our calibration dataset since https://unsloth.ai/docs/models/qwen3.5/gguf-benchmarks used a generic imatrix process from another provider for fairer comparisons.
After careful more experimentation, our dataset is still better in many cases. We plan to provide some metrics soon!
@danielhanchen I noticed in your Qwen3.5 benchmarks (https://unsloth.ai/docs/models/qwen3.5/gguf-benchmarks) that quantizing certain layers like attn_qkv has a negative impact. I see the new GGUFs for the 35B model (e.g., Qwen3.5-35B-A3B-UD-Q4_K_XL.gguf) have been updated to reflect this.
However, some of the current 397B quants, like Qwen3.5-397B-A17B-UD-Q4_K_XL, still have these layers quantized at MXFP4. Are you planning to apply the same layer updates to the 397B reuploads as you did for the 35B models?
*2026. March 3 2025 would be Qwen2.5.
Yeah, 2026, not 2025 :| ... fixed title
Hey folks please see https://www.reddit.com/r/LocalLLaMA/comments/1rlkptk/final_qwen35_unsloth_gguf_update/ which has all the details!