DeepSeek-V4.1-Flash-1017 Compact GGUF (for a 32 GB GPU + 96 GB RAM PC) â BETA
A general-purpose pruned build of deepseek-ai/DeepSeek-V4.1-Flash-1017 (MIT) - not code-only: 256 of 384 experts per layer, chosen by measured routing on chat, math, tool-call, code and multilingual text (kept routes: math 94%, tools 94%, English 91%, code 84%, Chinese 94%, Japanese 95%). 207 GB instead of the 276 GB of our unpruned LDLQ 132G, and faster on one 32 GB GPU plus a 96 GB RAM PC (87 vs 67.5 tok/s on coding tasks). Runs in Abyss and llama.cpp.
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
Why we made it: a code-only selection drops other abilities
ISTA-DASLab's GSQ-RCO Coder (192 of 384 experts, 182.7 GB) is small and fast, but in our tests it answered all three Japanese prompts in Chinese: its experts were selected on code, agentic, long-context and vision data, and our routing counts show that a code-only selection of 192 experts keeps only 45% of the routes Japanese text takes (code: 82%) - whatever the selection data leaves out is what goes first. JA256 selects its experts on Japanese, Chinese, English, code, math and tool-call text; at 256 experts it keeps 95% of the Japanese routes and 84% of the code routes (about the Coder's 82%). (ISTA's unpruned GSQ-RCO releases keep all experts; this is about the Coder only.)
Which one â this repository and its neighbours
| experts per layer | download | KL (8 seqs) | same top token | decode, 3 coding tasks | Japanese | |
|---|---|---|---|---|---|---|
JA256/ (here) |
256 / 384 | 207.1 GB | 0.383 | 75.2% | 87.0 tok/s | OK |
Full/ (here, planned) |
384 / 384 | ~230 GB | - | - | - | - |
LDLQ 132G |
384 / 384 | 275.8 GB | 0.194 | 81.1% | 67.5 tok/s | OK |
LDLQ 139G |
384 / 384 | 283.0 GB | 0.182 | 81.6% | 57.9 tok/s | OK |
| ISTA GSQ-RCO Coder (reference) | 192 / 384 | 182.7 GB | 0.582 | 69.3% | 97.0 tok/s | Chinese |
| ISTA GSQ-RCO IQ2_XXS (reference) | 384 / 384 | ~249 GB | 0.314 | 76.8% | - | - |
How to choose:
- JA256 â smaller download and the fastest file that keeps Japanese working: everyday chat, Japanese / Chinese / English, coding, math and tool calls. It is ~30% faster than 132G on our PC because its experts (88 GB) almost fit in GPU + RAM. Its KL is higher (about the same as ISTA's unpruned IQ1_M, 0.385), so long, careful work drifts more.
- LDLQ 132G / 139G â no expert removed, closest to the original. Pick these for hard coding work, long reasoning, other languages (Korean, European languages, ...), rare knowledge, or image input, and when 276-283 GB is fine.
- Full (coming) â every expert kept at ~230 GB, between the two.
- ISTA Coder â English coding only, if every GB and token per second counts; it does not answer Japanese.
What JA256 gives up: 128 of 384 experts per layer are gone. Text in the six domains above loses 5-16% of its routes to the next-best experts; anything outside them (other languages, rare domains, image input - not in our counts) may lose more and is weaker. The next version will follow your reports (e.g. weighting code higher).
Files (JA256)
JA256/DeepSeek-V4.1-Flash-1017-JA256-0000N-of-00006.gguf, 6 shards; point your engine at -00001-of-00006.gguf.
| shards 1-5 (dense + experts) | shard 6 (Engram table) | total | |
|---|---|---|---|
| JA256 | 96.5 GB (experts 88.4 GB) | 110.6 GB (IQ4_NL) | 207.1 GB |
- The experts, routers and router biases are LDLQ 132G's, sliced byte for byte to the 256 kept experts per layer
(
expert_count= 256; nothing re-quantized). Which experts are kept differs per layer. - The Engram table was re-quantized to IQ4_NL from the FP8 table in unsloth's UD-Q4_K_XL (relative RMS error 0.075 vs ISTA's IQ4_NL table 0.076 on the same rows). It is read from the SSD a few rows at a time and never has to fit in RAM.
SHA256SUMSin the folder.
Quality
KL against the unquantized model (top-256 log-probs of DeepSeek-V4.1-Flash-1017, 8 held-out sequences x 1,023 tokens), Abyss engine, exact routing: see the table above (per-sequence top-1 agreement of JA256: 72-79%).
KL is a guide only; real work (Abyss server, everyday settings, temperature 1.0 / top_p 0.95):
| JA256 | LDLQ 132G | ISTA Coder | |
|---|---|---|---|
| 10 hard coding tasks (hidden tests), solved | 8/10 (t104 solved in 2 of 2 reruns; t046 failed 3 of 3) | 10/10 | 9/10 (t046 failed) |
| quality score of the solved tasks (graded) | 80.5 | 77.6 | 77.5 |
| 8K / 32K prompt, find the needle | OK / OK | OK / OK | OK / OK |
| 7-turn tool-call coding session, tests pass | yes, final report in Japanese | yes | yes, final report in Chinese |
| Japanese (3 prompts), no mixed scripts | OK | OK | Chinese (no kana, one loop) |
| loops / illegal memory accesses | 0 / 0 | 0 / 0 | 1 / 0 |
The one task JA256 keeps failing (t046, object lifetimes in a dependency-injection container) is also the one the Coder fails: a sign of what pruning costs on hard code. If coding is your main use, take 132G / 139G.
Speed
Abyss v1.3.101, one PC: RTX 5090 (32 GB) on OCuLink PCIe Gen4 x4, Ryzen AI Max+ 395
(LPDDR5X, Radeon 8060S iGPU computes part of the RAM-side experts), RAM tier capped at 60 GB, NVMe SSD. Everyday
settings: DSpark drafter (Lucebox Q2K-Q4K), STRATA_GLM_TAIL_SKIP=4, STRATA_GLM_TAIL_SKIP_DISK=1,
STRATA_GLM_ROUTE_AFFINITY_RAM=0.08.
| decode tok/s | JA256 | LDLQ 132G | ISTA Coder |
|---|---|---|---|
| 3 coding tasks | 87.0 | 67.5 | 97.0 |
| 10 hard tasks (mean) | 75.7 | 58.0 | 87.1 |
| 32K prompt: prompt processing | 1,833 | 865 | 1,509 |
| 32K prompt: decode after it | 63.1 | 43.4 | 76.8 |
JA256's experts (88 GB) nearly fit in VRAM + the 60 GB RAM tier: 0.1 SSD reads a token against ~1 for 132G.
Other GPUs are not measured with this file. With 64 GB RAM, JA256 should lose less than the 384-expert files, since its experts are 44 GB smaller - reports welcome.
Feedback (this is a BETA)
- Speed on your PC:
./abyss.sh --bench --deepin Abyss, postabyss-bench.txtwith this issue form (change the title to your GPU), with your RAM size and SSD. - Quality in real use: where it went wrong, switched language, looped or got weaker than 132G - especially coding and languages outside the six domains. A prompt that reproduces it helps most.
- What you would rather have: more code weight, a smaller one for 64 GB RAM, other GPUs.
How it was made
With trench-gguf (tools/prune_experts.py, tools/requant_table.py) and
Abyss's routing counter (STRATA_GLM_USAGE):
- Routing counts: LDLQ 132G (all 384 experts) run over ~45K tokens each of Japanese, Chinese, English, code, math and tool-call text (eaddario/imatrix-calibration, micro sets) plus Japanese answers, counting which experts each layer routes to.
- Selection: per layer the 256 experts with the most routes, every domain weighted the same, chosen on 2/3 of the text and checked on the other 1/3 (kept routes: ja 94.8%, zh 93.6%, en 90.6%, code 83.6%, math 93.8%, tools 93.5%).
- Slicing:
prune_experts.pycuts the experts, router rows and biases of LDLQ 132G. - Engram:
requant_table.pyre-quantizes the FP8 table of unsloth UD-Q4_K_XL to IQ4_NL on the GPU (11 min).
Running it
- Abyss v1.3.101 or later. Pack it once:
and usecd Abyss .venv/bin/python tools/iq_pack.py --compat-bf16 --gguf /path/to/JA256/DeepSeek-V4.1-Flash-1017-JA256-00001-of-00006.gguf --out /path/to/JA256/packdocs/deepseek41/README.md's config with--glm-pack /path/to/JA256/pack. - llama.cpp (arch
deepseek41, upstreamac8e1694dof 2026-10-29): runsJA256as is (the expert count is read from the file). Checked withllama-server -m DeepSeek-V4.1-Flash-1017-JA256-00001-of-00006.gguf -ngl 99 --n-cpu-moe 36 -c 4096 --jinjaon the RTX 5090 (32 GB): correct answers to a Japanese question and a Python task, 15-20 tok/s (132G: 10-11 tok/s at--n-cpu-moe 38).
License
MIT, as the original model. All credit for the model goes to DeepSeek; the non-expert tensors come unchanged from unsloth/DeepSeek-V4.1-Flash-1017-GGUF.
æ¥æ¬èª
deepseek-ai/DeepSeek-V4.1-Flash-1017ïŒMITïŒã® GGUF ãã èªäœã® LDLQ 132G / 139G ããå°ããããŠçœ®ããŠããŸãã GPU 32 GB + RAM 96 GB ã® PC åãã§ã276 GB ã®ããŠã³ããŒãã¯éããšããæ¹ã®ããã®çã§ãã
BETA çã§ãã æž¬ã£ãã®ã¯äžã® 1 å°ã ãã§ããã»ãã® GPU ã RAM 容éã§ã®çµæããCommunity ã¿ãã Abyss ã® issue ã§æããŠãããããšå©ãããŸãã
äœã£ãçç±
ISTA-DASLab ã® GSQ-RCO Coder ïŒ384 åäž 192 åã182.7 GBïŒã¯å°ãããŠéãã®ã§ãããæå ã®ãã¹ãã§ã¯æ¥æ¬èªã®è³ªå 3 æ¬ãã¹ãŠã«äžåœèªã§çããŸããã ãšãã¹ããŒããã³ãŒãã»ãšãŒãžã§ã³ãã»é·æã»ç»åã®ããŒã¿ã§éžãã§ããããã§ãããã¡ãã®éèšã§ããã³ãŒãã ãã§ 192 åãéžã¶ãšã æ¥æ¬èªã®ã«ãŒã㯠45% ããæ®ããŸããïŒã³ãŒã㯠82%ïŒã JA256 ã¯æ¥æ¬èªã»äžåœèªã»è±èªã»ã³ãŒãã»æ°åŠã»ããŒã«åŒã³åºãã®æç« ã§éžã³ã256 åã§æ¥æ¬èªã®ã«ãŒãã® 95%ãã³ãŒãã® 84% ïŒCoder ãšã»ãŒåãïŒãæ®ããŠããŸããïŒISTA ã®ãšãã¹ããŒããåã£ãŠããªã GSQ-RCO çã¯å šéšæ®ããŠããã®ã§ããã®è©±ã¯ Coder çã ãã®ããšã§ããïŒ
䜿ãåã
| 1 å±€ã®ãšãã¹ããŒã | ããŠã³ããŒã | KLïŒ8 æ¬ïŒ | 1 äœã®äžèŽ | çæé床ïŒã³ãŒãã£ã³ã° 3 åïŒ | æ¥æ¬èª | |
|---|---|---|---|---|---|---|
JA256/ïŒããïŒ |
256 / 384 | 207.1 GB | 0.383 | 75.2% | 87.0 tok/s | OK |
Full/ïŒãããäºå®ïŒ |
384 / 384 | çŽ 230 GB | - | - | - | - |
LDLQ 132G |
384 / 384 | 275.8 GB | 0.194 | 81.1% | 67.5 tok/s | OK |
LDLQ 139G |
384 / 384 | 283.0 GB | 0.182 | 81.6% | 57.9 tok/s | OK |
| ISTA GSQ-RCO CoderïŒåèïŒ | 192 / 384 | 182.7 GB | 0.582 | 69.3% | 97.0 tok/s | äžåœèª |
| ISTA GSQ-RCO IQ2_XXSïŒåèïŒ | 384 / 384 | çŽ 249 GB | 0.314 | 76.8% | - | - |
- JA256 ãããããã®äººïŒããŠã³ããŒããå°ããããããéããã»ãããæ¥æ¬èªã䜿ãã
- æ®æ®µã®äŒè©±ãæ¥æ¬èªã»äžåœèªã»è±èªãã³ãŒãã£ã³ã°ãæ°åŠãããŒã«åŒã³åºãã«åããŸãã
- ãšãã¹ããŒãïŒ88 GBïŒã GPU + RAM ã«ã»ãŒåãŸãã®ã§ãæå ã® PC ã§ã¯ 132G ããçŽ 3 å²éãåããŸãã
- ãã®ããã KL ã¯å€§ããã§ãïŒISTA ã®åã£ãŠããªã IQ1_M ã® 0.385 ãšåããããïŒãé·ãæ éãªäœæ¥ã§ã¯ãå ã®ã¢ãã«ããããããããªããŸãã
- LDLQ 132G / 139G ãããããã®äººïŒãšãã¹ããŒãã 1 ã€ãåã£ãŠããããå
ã®ã¢ãã«ã«äžçªè¿ãçã§ãã
- é£ããã³ãŒãã£ã³ã°ãé·ãæšè«ãã»ãã®èšèªïŒéåœèªã欧å·ã®èšèªãªã©ïŒãçããç¥èãç»åå ¥åã䜿ãå Žåã
- 276ã283 GB ã®ããŠã³ããŒããèŠã«ãªããªãå Žåã
- FullïŒæºåäžïŒïŒãšãã¹ããŒããå šéšæ®ããŠçŽ 230 GB ã«ããçã§ãäžã® 2 ã€ã®äžéã§ãã
- ISTA CoderïŒè±èªã§ã³ãŒãã£ã³ã°ã ãã1 GB ã§ãå°ããéããããå Žåãæ¥æ¬èªã«ã¯çããããŸããã
JA256 ã§åŒ±ããªããã®ïŒåå±€ 384 åã®ãã¡ 128 åãåã£ãŠããŸããäžã® 6 ã€ã®é åã®æç« ã§ããã«ãŒãã® 5ã16% ããæ¬¡ã«è¿ãããšãã¹ããŒãã«åããŸãã éèšã«å ¥ããŠããªããã®ïŒã»ãã®èšèªãçããåéãç»åå ¥åïŒã¯ããã£ãšåŒ±ããªã£ãŠããå¯èœæ§ããããŸãã æ¬¡ã®çã¯ãããã ããå ±åãããšã«äœããŸãïŒããšãã°ã³ãŒãã®éã¿ãäžãããªã©ïŒã
äžèº«
JA256/㯠6 ãã¡ã€ã«ã§ 1 ã»ããã§ãã-00001-of-00006.ggufãæå®ããŠãã ãããåèš 207.1 GBïŒ1ã5 çªç® 96.5 GBããã®ãã¡ãšãã¹ããŒã 88.4 GBã6 çªç®ã® Engram 衚 110.6 GBïŒã- ãšãã¹ããŒãã»ã«ãŒã¿ãŒã»ã«ãŒã¿ãŒã® bias ã¯ãLDLQ 132G ã®ãã®ããæ®ã 256 åã®åã ããã€ãåäœã§åãåºããŠããŸãïŒéååã¯ããçŽããŠããŸãããæ®ããšãã¹ããŒãã¯å±€ããšã«éããŸãïŒã
- Engram 衚ã¯ãunsloth ã® UD-Q4_K_XL ã«å ¥ã£ãŠãã FP8 ã®è¡šãã IQ4_NL ã«äœãçŽããŸãããå ãšã®èª€å·®ã¯ 0.075 ã§ãISTA ã® IQ4_NLïŒ0.076ïŒãšåããããã§ããSSD ããå°ããã€èªãã®ã§ãRAM ã«èŒããå¿ èŠã¯ãããŸããã
å質ãšå®çšã»é床
| JA256 | LDLQ 132G | ISTA Coder | |
|---|---|---|---|
| é£ããã®ã³ãŒãã£ã³ã° 10 åïŒé ããã¹ãïŒ | 8/10ïŒt104 ã¯è§£ãçŽã 2 åãšãæ£è§£ãt046 㯠3 åãšãäžæ£è§£ïŒ | 10/10 | 9/10ïŒt046 ãèœãšãïŒ |
| è§£ããåé¡ã®åè³ªç¹ | 80.5 | 77.6 | 77.5 |
| 8K / 32K ã®é·æããæ å ±ãæ¢ã | æå / æå | æå / æå | æå / æå |
| ããŒã«ã䜿ã 7 åŸåŸ©ã®äœæ¥ | ãã¹ãééãæåŸã®å ±åã¯æ¥æ¬èª | éé | ééãå ±åã¯äžåœèª |
| æ¥æ¬èª 3 å | èªç¶ãªæ¥æ¬èª | èªç¶ | äžåœèªïŒã㪠0ã1 æ¬ã«ãŒãïŒ |
| 10 åã®å¹³åçæé床 | 75.7 tok/s | 58.0 tok/s | 87.1 tok/s |
| 32K ã®å ¥ååŠç | 1,833 tok/s | 865 tok/s | 1,509 tok/s |
- JA256 ãæ¯åèœãšãã 1 åïŒt046ãDI ã³ã³ããã®ãªããžã§ã¯ãã®å¯¿åœïŒã¯ãCoder ãèœãšããåé¡ã§ãããšãã¹ããŒããåã£ã圱é¿ããé£ããã³ãŒãã«åºãäŸãšèããŠããŸããã³ãŒãã£ã³ã°ãäž»ãªçšéãªãã132G / 139G ãããããããŸãã
- ãšãã¹ããŒãïŒ88 GBïŒã VRAM ãš RAM 60 GB ã«ã»ãŒåãŸãã®ã§ãSSD ããã®èªã¿åºã㯠1 ããŒã¯ã³ããã 0.1 åã§ãïŒ132G ã¯çŽ 1 åïŒã
- 枬å®ç°å¢ïŒRTX 5090ïŒ32 GBãOCuLink Gen4 x4ïŒãRyzen AI Max+ 395ãRAM ã®æ 60 GBãNVMe SSDãDSpark ãªã©ã®æ®æ®µäœ¿ãã®èšå®ã
- ã»ãã® GPU ã§ã¯ãŸã 枬ã£ãŠããŸããããšãã¹ããŒãã 44 GB å°ããã®ã§ãRAM 64 GB ã® PC ã§ã¯ã384 åã®çããéãã®èœã¡æ¹ãå°ããã¯ãã§ãã
ãã£ãŒãããã¯ã®ãé¡ãïŒBETA çïŒ
- ã䜿ãã® PC ã§ã®éãïŒAbyss ã§
./abyss.sh --bench --deepãå®è¡ããabyss-bench.txtã ãã¡ãã® issue ãã©ãŒã ã«è²Œã£ãŠãã ããïŒã¿ã€ãã«ã® GPU åãæžãæããŠãã ããïŒãRAM ã®å®¹éãš SSD ã®æ©çš®ãæžããŠãããããšå©ãããŸãã - å®éã«äœ¿ã£ããšãã®å質ïŒããããªçããèšèªã®åãæ¿ãããç¹°ãè¿ãã132G ãã匱ããšæãããšãããç¹ã«ã³ãŒãã£ã³ã°ãšã6 ã€ã®é å以å€ã®èšèªã®äŸãå©ãããŸããåçŸã§ããããã³ããããããšäžçªå©ãããŸãã
- 欲ããçïŒã³ãŒãå¯ãã®çãRAM 64 GB åãã®å°ããçãã»ãã® GPU åããªã©ã
äœãæ¹ã»åããæ¹ã»ã©ã€ã»ã³ã¹
- äœãæ¹ãšåããæ¹ã¯ãäžã®è±èªã®ç¯ã®ãšããã§ãïŒtrench-gguf ã®
prune_experts.pyãšrequant_table.pyãAbyss ã®ã«ãŒãéèšïŒã - Abyss v1.3.101 以éã§åäœã確èªããŠããŸããllama.cppïŒ2026-10-29 ã® ac8e1694dïŒã§ããã®ãŸãŸåããŸãããRTX 5090 ã§
--n-cpu-moe 36ã®ãšããæ¥æ¬èªã®è³ªåãš Python ã®èª²é¡ã«æ£ããçããçæã¯ 15ã20 tok/s ã§ããïŒ132G ã¯--n-cpu-moe 38ã§ 10ã11 tok/sïŒã - ã©ã€ã»ã³ã¹ã¯å ã®ã¢ãã«ãšåã MIT ã§ãã