DeepSeek-V4.1-Flash-1017 LDLQ GGUF (for a 32 GB GPU + 96 GB RAM PC) â BETA
Two GGUFs of deepseek-ai/DeepSeek-V4.1-Flash-1017 (MIT), made for one 32 GB GPU plus a 96 GB RAM PC. All 384 routed experts of every layer are kept (no expert pruning); the experts were re-quantized from the original FP4 weights with LDLQ (error-feedback rounding with each expert's own input statistics), so they are rounded better than plain imatrix quantization at the same types.
BETA: measured on one PC only (below). Reports from other GPUs / RAM sizes are very welcome (Community tab, or Abyss issues).
æ¥æ¬èªã®èª¬æã¯äžã«ãããŸãã
Which one
| folder | experts | for | KL (8 seqs) | hard tasks quality | decode, coding tasks |
|---|---|---|---|---|---|
132G/ |
132.2 GB | speed - same expert bytes as unsloth UD-IQ1_M | 0.194 | 77.6 | 67.5 tok/s |
139G/ |
139.4 GB | quality - layers 35-39 one step higher (IQ3_XXS) | 0.182 | 81.4 | 57.9 tok/s |
| unsloth UD-IQ1_M (reference, not here) | 132.2 GB | 0.198 | 81.1 | 71.8 tok/s |
- Pick 132G if you want the speed of UD-IQ1_M with slightly better numbers, 139G if you can give up 15-20% decode speed for the best quality at this size. Both pass every real-use check below.
- "experts" is the routed-expert bytes (decimal GB), the part that decides whether the model fits GPU + RAM.
Files
Each folder is a complete set of 6 shards; point your engine at -00001-of-00006.gguf.
| shards 1-5 (dense + experts) | shard 6 (Engram table) | total | |
|---|---|---|---|
132G/DeepSeek-V4.1-Flash-1017-LDLQ-132G-*.gguf |
140.6 GB | 135.2 GB | 275.8 GB |
139G/DeepSeek-V4.1-Flash-1017-LDLQ-139G-*.gguf |
147.8 GB | 135.2 GB | 283.0 GB |
- Shard 6 is the Engram table (Q5_K), byte-identical to unsloth's
UD-IQ2_XXS-00006-of-00006.gguf. It is read from the SSD a few rows at a time and never has to fit in RAM. - Everything except the routed experts (attention, shared experts, routers, embeddings, output, metadata, tokenizer) is byte-identical to unsloth's UD-IQ2_XXS (dense Q8_0).
SHA256SUMSin each folder.
What is different from unsloth's files
| layers | gate / up | down | 139G: layers 35-39 |
|---|---|---|---|
| 0-4 | IQ1_S | IQ1_S | |
| 5-19 | IQ1_S | IQ2_XXS | |
| 20-29 | IQ2_XXS | IQ2_S | |
| 30-37 | IQ2_XXS | IQ2_XS | 35-37: all IQ3_XXS |
| 38-39 | IQ2_XS | IQ2_XS | all IQ3_XXS |
The types of 132G are exactly unsloth UD-IQ1_M's; only the rounding differs. Plain re-quantization at these types
was far worse (KL 0.404, 8 sequences; the errors of the last five layers add up), unsloth's own rounding gets 0.198, and LDLQ gets
0.194. In 139G the last five layers, where the errors pile up, are one step higher.
Quality
KL against the unquantized model (top-256 log-probs of DeepSeek-V4.1-Flash-1017 before quantization, 8 held-out sequences x 1,023 tokens), computed in the Abyss engine with exact routing.
| experts | mean KL | same top token | nll | |
|---|---|---|---|---|
| unsloth UD-IQ2_XXS (rev 85f81fb) | 156.9 GB | 0.142 | 83.7% | 2.173 |
| 139G | 139.4 GB | 0.182 | 81.6% | 2.204 |
| 132G | 132.2 GB | 0.194 | 81.1% | 2.215 |
| unsloth UD-IQ1_M | 132.2 GB | 0.198 | 81.7% | 2.209 |
KL is a guide only, so both were also run on real work (Abyss server, everyday settings below, temperature 1.0 / top_p 0.95):
| 132G | 139G | UD-IQ1_M | UD-IQ2_XXS | |
|---|---|---|---|---|
| 10 hard coding tasks (hidden tests), solved | 10/10 | 10/10 | 10/10 | 10/10 |
| quality score (graded) | 77.6 | 81.4 | 81.1 | 78.8 |
| 8K / 32K prompt, find the needle | OK / OK | OK / OK | OK / OK | OK / OK |
| 7-turn tool-call coding session, tests pass | yes | yes | yes | yes |
| Japanese (3 prompts), no mixed scripts | OK | OK | OK | OK |
132G's lower quality score comes from one task (47 points); the other nine are level with the rest.
Speed
Abyss v1.3.9x, one PC: RTX 5090 (32 GB) on OCuLink PCIe Gen4 x4, Ryzen AI Max+ 395
(16 cores, LPDDR5X; its Radeon 8060S iGPU computes part of the RAM-side experts), RAM tier capped at 60 GB
(STRATA_GLM_RAM_GB=60), the rest read from an NVMe SSD. Everyday settings: DSpark drafter (Lucebox's DeepSeek-V4.1-Flash-1017 DSpark draft, Q2K-Q4K), STRATA_GLM_TAIL_SKIP=4, STRATA_GLM_TAIL_SKIP_DISK=1,
STRATA_GLM_ROUTE_AFFINITY_RAM=0.08.
| decode tok/s | 132G | 139G | UD-IQ1_M |
|---|---|---|---|
| 3 coding tasks (same day, one run each) | 67.5 | 57.9 | 71.8 |
| 10 hard tasks (mean) | 58.0 | 47.8 | 52.5 |
| 32K prompt, then decode | 43.4 | 43.4 | - |
132G and UD-IQ1_M swap places from run to run (the two sets disagree by ±5%); 139G is clearly slower on this PC because 7 GB more has to come from RAM / SSD. A desktop with dual-channel DDR5 and no iGPU will be slower than these numbers.
Other GPUs (rough guide)
Only the RTX 5090 row is measured with these files. The others are labelled for what they are:
| GPU (VRAM) | 132G | 139G | basis |
|---|---|---|---|
| RTX 5090 (32 GB) | 67.5 tok/s | 57.9 tok/s | measured (above) |
| Radeon AI PRO R9700 (32 GB) | ~40 tok/s | ~33-35 tok/s | estimate: unsloth UD-IQ1_M - the same types and bytes as 132G, only the rounding differs - measured 40.6 tok/s on an R9700 + 96 GB RAM (8 coding tasks, DSpark; 30.5 with 64 GB RAM). 139G scaled by its 5090 slowdown. To be replaced by measurements |
| V100 (32 GB) | not measured | not measured | Abyss builds for sm_70 with CUDA 12.9 (--nvcc) but has not run on a V100 yet; no FP4 / newer tensor-core kernels there, so expect less than the R9700. Reports welcome |
| RTX 3090 / 4090 (24 GB) | not measured | not measured | 8 GB under this repo's 32 GB target: more experts come from the SSD, so speed follows the SSD. Use 132G; reports welcome |
Runtime options
Abyss has runtime options that change how the model answers without changing these files (for example
--uncensor, see docs/UNCENSOR.md). The files here are
the normal model.
Feedback (this is a BETA)
The next versions of these files will follow what users report. Useful are:
- Speed on your PC: run
./abyss.sh --bench --deepin Abyss and postabyss-bench.txtwith this issue form (it already has the right template; change the title to your GPU). Add your RAM size and SSD. - Quality in real use: tasks where it went wrong, broke into another language or looped, or refused too much, in the Community tab here or an Abyss issue. A prompt that reproduces it helps most.
- What you would rather have: a smaller file for 64 GB RAM, a bigger one for 128 GB, other GPUs.
Measured reports go into the GPU table above with credit; problems that show up go into the next build.
How it was made
All with trench-gguf v0.3.0 (GPU quantizer, LDLQ). Every command and both rules files are in its README, LDLQ walkthrough:
- Source: unsloth's
UD-Q4_K_XLof DeepSeek-V4.1-Flash-1017, whose routed experts are the original MXFP4 (FP4) weights. - Hessians: every layer's expert inputs dumped from a forward pass over 491,520 tokens of general English / Japanese
text and code (
hessian_dump, a patched llama.cpp on the GPU, 49 min on the RTX 5090). - Re-quantization:
requant_experts.py --gpu --dump=<hessians> --order=act --damp=0.3 --damp-iq1=1.0with unsloth's imatrix, the type table above (132G), then layers 35-39 again at IQ3_XXS (139G). - Assembly:
merge_experts.pyputs the new expert tensors into unsloth UD-IQ2_XXS (rev 85f81fb) byte for byte; the result's Abyss pack was checked to have exactly the measured layout.
Experts with fewer than 64 calibration rows (a few per layer) fall back to plain imatrix rounding.
Running it
- Abyss v1.3.100 or later, one command (engine, download with SHA-256 check,
pack, config with the everyday settings below):
./setup-deepseek.sh --model v41 --download-model(--quant 139Gfor the quality build). By hand (tested): pack it once (the pack indexes the shards in place):
and usecd Abyss .venv/bin/python tools/iq_pack.py --compat-bf16 --gguf /path/to/132G/DeepSeek-V4.1-Flash-1017-LDLQ-132G-00001-of-00006.gguf --out /path/to/132G/packdocs/deepseek41/README.md's config with--glm-pack /path/to/132G/pack. - llama.cpp (arch
deepseek41, upstreamac8e1694dof 2026-10-29): runs132Gas is. Checked withllama-server -m DeepSeek-V4.1-Flash-1017-LDLQ-132G-00001-of-00006.gguf -ngl 99 --n-cpu-moe 38 -c 4096 --jinjaon the RTX 5090 (32 GB): correct answers to a Japanese question and a Python task, 10-11 tok/s (experts of two layers on the GPU, the rest from RAM / SSD).--n-cpu-moe 34ran out of VRAM on the first request.
License
MIT, as the original model. This repository only holds a quantized conversion; all credit for the model goes to DeepSeek, and the non-expert tensors and Engram table come unchanged from unsloth/DeepSeek-V4.1-Flash-1017-GGUF.
æ¥æ¬èª
deepseek-ai/DeepSeek-V4.1-Flash-1017ïŒMITïŒã® GGUF ã 2 çš®é¡çœ®ããŠããŸãã GPU 32 GB + RAM 96 GB ã® PC ã§å¿«é©ã«åããããšãç®æšã«äœããŸããããšãã¹ããŒã㯠1 ã€ãåã£ãŠããŸããã ã«ãŒãã£ã³ã°ããããšãã¹ããŒãã¯ãå ã® FP4 ã®éã¿ãã LDLQïŒãšãã¹ããŒãããšã®å ¥åã®çµ±èšã䜿ããäžžãã®èª€å·®ã次ã®åã«éãæ¹æ³ïŒã§äœãçŽããŠããŸãã ãã®ãããåãåã®æ®éã® imatrix éååããäžžãã®èª€å·®ãå°ãããªããŸãã
BETA çã§ãã æž¬ã£ãã®ã¯äžã® 1 å°ã ãã§ããã»ãã® GPU ã RAM 容éã§è©ŠããçµæããCommunity ã¿ãã Abyss ã® issue ã§æããŠãããããšå©ãããŸãã
ã©ã¡ããéžã¶ã
| ãã©ã«ã | ãšãã¹ããŒã | åã | KLïŒ8 æ¬ïŒ | é£ãã 10 åã®å質 | çæé床ïŒã³ãŒãã£ã³ã°èª²é¡ïŒ |
|---|---|---|---|---|---|
132G/ |
132.2 GB | éãéèŠïŒunsloth UD-IQ1_M ãšåã倧ããïŒ | 0.194 | 77.6 | 67.5 tok/s |
139G/ |
139.4 GB | å質éèŠïŒå±€ 35ã39 ã 1 段äžã® IQ3_XXS ã«ïŒ | 0.182 | 81.4 | 57.9 tok/s |
| unsloth UD-IQ1_MïŒæ¯èŒçšãããã«ã¯ãããŸããïŒ | 132.2 GB | 0.198 | 81.1 | 71.8 tok/s |
- UD-IQ1_M ãšåããããã®éãã§ãæ°åãå°ãè¯ããã®ã欲ãããªã 132G ãéžãã§ãã ããã
- çæã 15ã20% é ããªã£ãŠãããã®ã§ããã®å€§ããã§äžçªã®åè³ªãæ¬²ãããªã 139G ã§ãã
- ã©ã¡ãããäžã®å®çšãã§ãã¯ããã¹ãŠéã£ãŠããŸãã
äžèº«
- åãã©ã«ã㯠6 ãã¡ã€ã«ã§ 1 ã»ããã§ãã
-00001-of-00006.ggufãæå®ããŠãã ããã - 6 çªç®ã®ãã¡ã€ã«ã¯ Engram ã®è¡šã§ããunsloth ã®ãã®ãšãŸã£ããåãã§ãSSD ããå°ããã€èªãã®ã§ RAM ã«èŒããå¿ èŠã¯ãããŸããã
- ãšãã¹ããŒã以å€ïŒattentionãå ±æãšãã¹ããŒããã«ãŒã¿ãŒãåã蟌ã¿ãåºåãããŒã¯ãã€ã¶ãŒïŒã¯ãunsloth UD-IQ2_XXS ãšãŸã£ããåãã§ãïŒdense 㯠Q8_0ïŒã
- 132G ã®åã®é
眮㯠unsloth UD-IQ1_M ãšãŸã£ããåãã§ãéãã®ã¯äžžãæ¹ã ãã§ãã
- ãã®åã®ãŸãŸæ®éã«éååãçŽããšã倧ãã厩ããŸããïŒKL 0.404ã8 ç³»åïŒãæåŸã® 5 å±€ã®èª€å·®ãç©ã¿éãªãããã§ãã
- unsloth ç¬èªã®äžžãã§ã¯ 0.198ãLDLQ ã§ã¯ 0.194 ã«ãªããŸããã
- 139G ã¯ã誀差ãããŸããããæåŸã® 5 å±€ã 1 段äžã®åã«ããŠããŸãã
å質ãšå®çš
- KLïŒ éåååã®ã¢ãã«ãšã®å·®ãã8 æ¬ Ã 1,023 ããŒã¯ã³ã§æž¬ããŸããã
- unsloth UD-IQ2_XXSïŒ156.9 GBïŒïŒ0.142
- 139GïŒ0.182
- 132GïŒ0.194
- unsloth UD-IQ1_MïŒ0.198
- å®éã®äœæ¥ïŒ KL ã¯ãããŸã§ç®å®ãªã®ã§ãå®éã®äœæ¥ã§ã確ãããŸãããã©ã¡ããæ¬¡ã®ãã§ãã¯ããã¹ãŠéã£ãŠããŸãã
- é£ããã®ã³ãŒãã£ã³ã° 10 åïŒé ããã¹ãããïŒïŒå šåæ£è§£
- 8K / 32K ã®é·æããæå®ããæ å ±ãæ¢ãïŒæå
- ããŒã«ã䜿ã 7 åŸåŸ©ã®ã³ãŒãã£ã³ã°äœæ¥ïŒãã¹ãéé
- æ¥æ¬èª 3 åïŒæååãã»ä»èšèªã®æ··å ¥ãªã
- 132G ã®å質ç¹ãå°ãäœãã®ã¯ã1 åã ã 47 ç¹ã ã£ãããã§ããæ®ãã® 9 åã¯ã»ããšåãæ°Žæºã§ãã
é床
枬å®ç°å¢ïŒ RTX 5090ïŒ32 GBãOCuLink Gen4 x4ïŒãRyzen AI Max+ 395ïŒiGPU ã RAM åŽã®ãšãã¹ããŒãã®èšç®ã«äœ¿çšïŒãRAM ã®æ 㯠60 GBãæ®ã㯠NVMe SSD ããèªã¿åºããDSpark ã®äžæžããªã©ãæ®æ®µäœ¿ãã®èšå®ã§ãã
çæé床 132G 139G UD-IQ1_M ã³ãŒãã£ã³ã°èª²é¡ 3 å 67.5 tok/s 57.9 tok/s 71.8 tok/s é£ããã® 10 å 58.0 tok/s 47.8 tok/s 52.5 tok/s 132G ãš UD-IQ1_M ã¯æž¬ããã³ã«é äœãå ¥ãæ¿ããçšåºŠã®å·®ïŒÂ±5%ïŒã§ãã139G ã¯ãRAM ã SSD ããèªãéã 7 GB å¢ããåãã¯ã£ããé ããªããŸãã
DDR5 ãã¥ã¢ã«ãã£ãã«ã§ iGPU ã®ãªããã¹ã¯ãããã§ã¯ããã®æ°åããé ããªããŸãã
ã»ãã® GPU ã§ã®ç®å®
ãã®ãã¡ã€ã«ã§å®æž¬ããã®ã¯ RTX 5090 ã ãã§ããã»ãã¯ãæ ¹æ ãæžããããã§ç®å®ãšããŠèŒããŠããŸãã
| GPUïŒVRAMïŒ | 132G | 139G | æ ¹æ |
|---|---|---|---|
| RTX 5090ïŒ32 GBïŒ | 67.5 tok/s | 57.9 tok/s | 宿ž¬ïŒäžã®è¡šïŒ |
| Radeon AI PRO R9700ïŒ32 GBïŒ | çŽ 40 tok/s | çŽ 33ã35 tok/s | ç®å®ïŒ132G ãšåã»ãã€ãæ°ãåã unsloth UD-IQ1_MïŒéãã®ã¯äžžãã ãïŒã R9700 + RAM 96 GB ã§å®æž¬ãããš 40.6 tok/s ã§ããïŒã³ãŒãã£ã³ã° 8 åãDSpark ãããRAM 64 GB ã§ã¯ 30.5ïŒã139G 㯠5090 ã§ã®é ããªãæ¹ããæç®ããŠããŸããããšã§å®æž¬ã«çœ®ãæããŸã |
| V100ïŒ32 GBïŒ | æªæž¬å® | æªæž¬å® | Abyss 㯠CUDA 12.9ïŒ--nvccïŒã§ sm_70 åãã«ãã«ãã§ããŸãããV100 ã®å®æ©ã§ã¯ãŸã åãããŠããŸãããFP4 ãæ°ãã tensor core ã®ã«ãŒãã«ã䜿ããªãã®ã§ãR9700 ããé
ããšæãããŸããçµæãæããŠãããããšå©ãããŸã |
| RTX 3090 / 4090ïŒ24 GBïŒ | æªæž¬å® | æªæž¬å® | ãã®ãªããžããªãæ³å®ãã 32 GB ãã 8 GB å°ãªããããSSD ããèªããšãã¹ããŒããå¢ããéãã¯ã»ãŒ SSD 次第ã«ãªããŸãã䜿ããªã 132G ãããããã§ããçµæãæããŠãããããšå©ãããŸã |
å®è¡æã®ãªãã·ã§ã³
Abyss ã«ã¯ããã¡ã€ã«ãå€ããã«çãæ¹ãå€ããå®è¡æã®ãªãã·ã§ã³ããããŸãïŒããšãã° --uncensorã詳ãã㯠docs/UNCENSORïŒãããã«ãããã¡ã€ã«ã¯ãæ®éã®ã¢ãã«ã§ãã
ãã£ãŒãããã¯ã®ãé¡ãïŒBETA çïŒ
ãã®ãã¡ã€ã«ã®æ¬¡ã®çã¯ã䜿ã£ãæ¹ããã®å ±åãããšã«äœããŸããç¹ã«æ¬¡ã® 3 ã€ãå©ãããŸãã
- ã䜿ãã® PC ã§ã®éãïŒ
- Abyss ã§
./abyss.sh --bench --deepãå®è¡ããŠãabyss-bench.txtããã¡ãã® issue ãã©ãŒã ã«è²Œã£ãŠãã ããããã³ãã¬ãŒãã¯èšå®æžã¿ãªã®ã§ãã¿ã€ãã«ã® GPU åã ãæžãæããŠãã ããã - RAM ã®å®¹éãš SSD ã®æ©çš®ãæžããŠãããããšå©ãããŸãã
- Abyss ã§
- å®éã«äœ¿ã£ããšãã®å質ïŒ
- ããããªçãã«ãªã£ããå¥ã®èšèªãæ··ãã£ããåãæãç¹°ãè¿ããããæããããããšãã£ãäŸãæããŠãã ããã
- ããã® Community ã¿ãããAbyss ã® issue ãžãé¡ãããŸããåçŸã§ããããã³ããããããšäžçªå©ãããŸãã
- 欲ããçïŒ RAM 64 GB åãã®å°ããçã128 GB åãã®å€§ããçãã»ãã® GPU åããªã©ã
ããã ãã宿ž¬å€ã¯ãäžã® GPU ã®è¡šã«ååãæ·»ããŠèŒããŸããèŠã€ãã£ãåé¡ã¯ã次ã®çã§çŽããŸãã
äœãæ¹
- éå ·ã¯ãã¹ãŠ trench-gguf v0.3.0ïŒGPU ã§éååããéå ·ãLDLQ 察å¿ïŒã§ããã³ãã³ããšèŠåãã¡ã€ã«ã¯ãREADME ã® LDLQ ã®å®äŸ ã«ãã¹ãŠèŒã£ãŠããŸãã
- æé ïŒ
- å ã®éã¿ã«ã¯ãunsloth ã® UD-Q4_K_XL ã䜿ããŸãããããã«å ¥ã£ãŠãããšãã¹ããŒãã¯ãå ã® MXFP4ïŒFP4ïŒã®ãŸãŸã§ãã
- äžè¬çãªè±èªã»æ¥æ¬èªã»ã³ãŒãã®æç« 491,520 ããŒã¯ã³ãæµããŠãåå±€ã®ãšãã¹ããŒããžã®å ¥åãåãåºããŸããïŒRTX 5090 ã§ 49 åïŒã
- ããã䜿ããLDLQ ã§ãšãã¹ããŒããäœãçŽããŸãããimatrix 㯠unsloth ã®ãã®ã§ãã
- äœãçŽãããšãã¹ããŒãããunsloth UD-IQ2_XXS ã« 1 ãã€ããå€ããã«å·®ã蟌ã¿ãŸããã
åããæ¹ã»ã©ã€ã»ã³ã¹
- Abyss v1.3.100 以éãªãã
./setup-deepseek.sh --model v41 --download-modelã® 1 è¡ã§ããšã³ãžã³ã®ãã«ããããŠã³ããŒãïŒSHA-256 ã®ç §åã€ãïŒãpackãæ®æ®µäœ¿ãã®èšå®ã®å ¥ã£ãèšå®ãã¡ã€ã«ãŸã§æžã¿ãŸããå質éèŠã®çã¯--quant 139Gãè¶³ããŠãã ããã - æäœæ¥ã®åããæ¹ãšã©ã€ã»ã³ã¹ã¯ãäžã®è±èªã®ç¯ã®ãšããã§ãïŒAbyss ã§åäœç¢ºèªæžã¿ïŒã
- llama.cppïŒ2026-10-29 ã® ac8e1694dïŒã§ã
132Gããã®ãŸãŸåããŸãããRTX 5090 ã§--n-cpu-moe 38ã®ãšããæ¥æ¬èªã®è³ªåãš Python ã®èª²é¡ã«æ£ããçããçæã¯ 10ã11 tok/s ã§ããïŒ--n-cpu-moe 34ã§ã¯æåã®ãªã¯ãšã¹ãã§ VRAM ãè¶³ããªããªããŸããïŒã - ã©ã€ã»ã³ã¹ã¯å ã®ã¢ãã«ãšåã MIT ã§ãã