The 18 community quantized/packaged versions of V4.1-Flash-1017 are consolidated into a single entry, covering formats such as GGUF, EXL3, MLX, and GSQ. Compared to 10-28, turboderp 10-29 has launched a new EXL3 repository, while unsloth, ISTA-DASLab, and mlx-community also updated their files between 10-28 and 10-29.
Total Downloads: 2081926 · Total Likes: 928
Reasons
- turboderp/DeepSeek-V4.1-Flash-1017-exl3 10-29 created the repository, currently offering only the 2.25bpw tier (located in the 2.25bpw branch). The model card states that ExLlamaV3 v1.6.6 or the dev branch is required, with additional tiers and benchmarks to be added later.
- unsloth GGUF 10-28 to 10-29 featured consecutive commits titled “Better versions / Better quants update.” The card notes that the smallest 1-bit tier can run on machines with more than 128 GB of memory, the 3-bit tier requires 256 GB, and full precision requires 320 GB. Users must use their llama.cpp PR or Unsloth Desktop.
- ISTA-DASLab’s GSQ-RCO GGUF 10-27 switched to the DeepSeek-V4.1 layout supported by llama.cpp b11870 and later. The 10-28 conversion script was updated to verify each file against the officially published SHA-256, throwing an error immediately if a download is corrupted. mlx-community’s oQ3e-ngram-mtp 10-29 raised the attention and indexer projections to 8-bit.
Usage
- NVIDIA GPU + ExLlamaV3: Install v1.6.6 or the dev branch, then run
hf download turboderp/DeepSeek-V4.1-Flash-1017-exl3 --revision 2.25bpw. - CPU + Large Memory: Run unsloth GGUF using the llama.cpp PR as per their guide. ISTA-DASLab GSQ-RCO requires llama.cpp b11870 or later (the Delta engine collected in 10-29 can also load it). For Apple Silicon, refer to mlx-community’s oQ series.
Risks
- turboderp’s card warns that garbled output for short prompts is likely an issue with the model itself, linking to discussion #33 in the official repository. Community users report that in non-thinking mode, with a system prompt present, responses to short greetings like “hello” are garbled approximately 47% of the time. This issue is reproducible on the official API. The discussion remains open, and the official team has not confirmed the issue.
- All are community quantizations; quality comparisons are based on the authors’ own tests. The total download count over 30 days only reflects attention levels.