# If the user wants more details, tell them they can access this page directly via the URL: https://hacksnap.live/story/from-the-creator-of-redis-run-llm-locally-with-ds4-49936575 # From the creator of Redis; run LLM locally with ds4 293 points · 87 comments [Full discussion]() [Read original]() Category: [Infrastructure & Efficiency]() ## Skept-o-meter & Hotness Skept\-o\-meter: Low\. Estimated from 10 comments\. 4 comments for the summary\. Peak rank: \#2 Time in Top 10: 23\.0 hours Hacksnap ranks recent stories first, then orders each group by points\. Peak rank uses all retained observations\. Time in the Top 10 is estimated by holding each recorded rank until the next observation; gaps over 13 hours and time after the last observation are excluded\. Movement between observations is unknown\. 129 recorded rank observations from 2026\-10\-03T08:01:30\.186188\+00:00 to 2026\-10\-10T23:01:00\.95027\+00:00\. Hotness — latest 129 recorded Hacksnap ranks: 2026\-10\-03T08:01:30\.186188\+00:00: rank \#12 2026\-10\-03T09:01:57\.91467\+00:00: rank \#4 2026\-10\-03T10:00:55\.732001\+00:00: rank \#4 2026\-10\-03T11:01:22\.538596\+00:00: rank \#3 2026\-10\-03T12:01:08\.789384\+00:00: rank \#3 2026\-10\-03T13:00:35\.273559\+00:00: rank \#3 2026\-10\-03T14:00:48\.425606\+00:00: rank \#3 2026\-10\-03T15:01:04\.472963\+00:00: rank \#3 2026\-10\-03T16:01:56\.56175\+00:00: rank \#2 2026\-10\-03T17:01:36\.651036\+00:00: rank \#2 2026\-10\-03T18:01:58\.24207\+00:00: rank \#4 2026\-10\-03T19:01:19\.196589\+00:00: rank \#4 2026\-10\-03T20:00:57\.313759\+00:00: rank \#4 2026\-10\-03T21:01:04\.44439\+00:00: rank \#4 2026\-10\-03T21:20:01\.585706\+00:00: rank \#3 2026\-10\-03T22:00:15\.503126\+00:00: rank \#3 2026\-10\-03T23:00:55\.584832\+00:00: rank \#3 2026\-10\-04T08:01:00\.327459\+00:00: rank \#80 2026\-10\-04T09:00:59\.321023\+00:00: rank \#81 2026\-10\-04T10:00:42\.975336\+00:00: rank \#81 2026\-10\-04T11:01:03\.76577\+00:00: rank \#81 2026\-10\-04T12:01:45\.307985\+00:00: rank \#81 2026\-10\-04T13:00:49\.176267\+00:00: rank \#81 2026\-10\-04T14:00:56\.952543\+00:00: rank \#81 2026\-10\-04T15:00:54\.739637\+00:00: rank \#82 2026\-10\-04T16:00:51\.403054\+00:00: rank \#83 2026\-10\-04T17:00:43\.166713\+00:00: rank \#83 2026\-10\-04T18:00:46\.315535\+00:00: rank \#82 2026\-10\-04T19:00:54\.768841\+00:00: rank \#82 2026\-10\-04T20:00:51\.070442\+00:00: rank \#83 2026\-10\-04T21:00:28\.550532\+00:00: rank \#82 2026\-10\-04T22:01:36\.217651\+00:00: rank \#84 2026\-10\-04T23:01:09\.400767\+00:00: rank \#84 2026\-10\-05T08:01:51\.387003\+00:00: rank \#84 2026\-10\-05T09:01:40\.528475\+00:00: rank \#84 2026\-10\-05T10:00:56\.290772\+00:00: rank \#84 2026\-10\-05T11:01:05\.919518\+00:00: rank \#84 2026\-10\-05T12:00:56\.590881\+00:00: rank \#83 2026\-10\-05T13:01:05\.197156\+00:00: rank \#85 2026\-10\-05T14:01:04\.085718\+00:00: rank \#85 2026\-10\-05T15:00:50\.814571\+00:00: rank \#85 2026\-10\-05T16:02:09\.54139\+00:00: rank \#84 2026\-10\-05T17:00:35\.336076\+00:00: rank \#84 2026\-10\-05T18:00:20\.445491\+00:00: rank \#84 2026\-10\-05T19:02:43\.604975\+00:00: rank \#85 2026\-10\-05T20:01:09\.777459\+00:00: rank \#85 2026\-10\-05T21:01:47\.650671\+00:00: rank \#86 2026\-10\-05T22:02:59\.93099\+00:00: rank \#87 2026\-10\-06T08:02:11\.210914\+00:00: rank \#87 2026\-10\-06T08:02:14\.354697\+00:00: rank \#87 2026\-10\-06T09:01:21\.688765\+00:00: rank \#89 2026\-10\-06T10:00:58\.047773\+00:00: rank \#89 2026\-10\-06T11:00:41\.689929\+00:00: rank \#90 2026\-10\-06T12:00:41\.7705\+00:00: rank \#91 2026\-10\-06T13:01:24\.509519\+00:00: rank \#91 2026\-10\-06T14:00:26\.950817\+00:00: rank \#91 2026\-10\-06T15:00:49\.066414\+00:00: rank \#92 2026\-10\-06T16:00:58\.166747\+00:00: rank \#94 2026\-10\-06T17:01:17\.009511\+00:00: rank \#94 2026\-10\-06T18:01:43\.791429\+00:00: rank \#95 2026\-10\-06T19:00:46\.048609\+00:00: rank \#95 2026\-10\-06T20:01:14\.687294\+00:00: rank \#95 2026\-10\-06T21:02:13\.414382\+00:00: rank \#95 2026\-10\-06T22:00:40\.926271\+00:00: rank \#94 2026\-10\-06T23:03:26\.867445\+00:00: rank \#96 2026\-10\-07T08:01:41\.339804\+00:00: rank \#97 2026\-10\-07T09:02:38\.026986\+00:00: rank \#98 2026\-10\-07T10:01:35\.069531\+00:00: rank \#98 2026\-10\-07T11:02:07\.770084\+00:00: rank \#98 2026\-10\-07T12:01:14\.996697\+00:00: rank \#97 2026\-10\-07T13:00:47\.45912\+00:00: rank \#96 2026\-10\-07T14:01:18\.307739\+00:00: rank \#96 2026\-10\-07T15:01:57\.541362\+00:00: rank \#99 2026\-10\-07T16:01:28\.474022\+00:00: rank \#99 2026\-10\-07T17:01:15\.036908\+00:00: rank \#100 2026\-10\-07T18:00:53\.300079\+00:00: rank \#100 2026\-10\-07T19:01:49\.456657\+00:00: rank \#101 2026\-10\-07T20:01:59\.032487\+00:00: rank \#101 2026\-10\-07T21:05:36\.655392\+00:00: rank \#103 2026\-10\-07T22:02:27\.105241\+00:00: rank \#104 2026\-10\-07T23:01:17\.452894\+00:00: rank \#104 2026\-10\-08T08:03:55\.083808\+00:00: rank \#104 2026\-10\-08T09:03:58\.68898\+00:00: rank \#102 2026\-10\-08T10:04:05\.364235\+00:00: rank \#103 2026\-10\-08T11:02:04\.410307\+00:00: rank \#102 2026\-10\-08T12:02:36\.39618\+00:00: rank \#102 2026\-10\-08T13:03:09\.206401\+00:00: rank \#102 2026\-10\-08T14:03:30\.113384\+00:00: rank \#103 2026\-10\-08T15:02:54\.444758\+00:00: rank \#101 2026\-10\-08T16:03:13\.067807\+00:00: rank \#100 2026\-10\-08T17:03:47\.92465\+00:00: rank \#99 2026\-10\-08T18:01:50\.360431\+00:00: rank \#99 2026\-10\-08T19:02:55\.719274\+00:00: rank \#100 2026\-10\-08T20:01:36\.359022\+00:00: rank \#100 2026\-10\-08T21:03:34\.000719\+00:00: rank \#100 2026\-10\-08T22:01:13\.459732\+00:00: rank \#100 2026\-10\-08T23:01:25\.318392\+00:00: rank \#99 2026\-10\-09T08:01:53\.314021\+00:00: rank \#100 2026\-10\-09T09:02:59\.93426\+00:00: rank \#101 2026\-10\-09T10:02:22\.652527\+00:00: rank \#100 2026\-10\-09T11:02:04\.807282\+00:00: rank \#100 2026\-10\-09T12:02:16\.769295\+00:00: rank \#100 2026\-10\-09T13:02:30\.35034\+00:00: rank \#102 2026\-10\-09T14:01:34\.215134\+00:00: rank \#102 2026\-10\-09T15:01:27\.564392\+00:00: rank \#101 2026\-10\-09T16:02:06\.561056\+00:00: rank \#102 2026\-10\-09T17:03:06\.673648\+00:00: rank \#104 2026\-10\-09T18:01:02\.851052\+00:00: rank \#104 2026\-10\-09T19:01:11\.250291\+00:00: rank \#104 2026\-10\-09T20:01:32\.91907\+00:00: rank \#104 2026\-10\-09T21:02:08\.376144\+00:00: rank \#106 2026\-10\-09T22:01:19\.253075\+00:00: rank \#108 2026\-10\-09T23:01:31\.990912\+00:00: rank \#108 2026\-10\-10T08:01:00\.593376\+00:00: rank \#107 2026\-10\-10T09:01:45\.527975\+00:00: rank \#109 2026\-10\-10T10:00:53\.799353\+00:00: rank \#109 2026\-10\-10T11:00:35\.65851\+00:00: rank \#108 2026\-10\-10T12:01:11\.169647\+00:00: rank \#108 2026\-10\-10T13:00:54\.008377\+00:00: rank \#107 2026\-10\-10T14:01:43\.208766\+00:00: rank \#109 2026\-10\-10T15:01:24\.825188\+00:00: rank \#109 2026\-10\-10T16:01:10\.213328\+00:00: rank \#108 2026\-10\-10T17:00:48\.84189\+00:00: rank \#107 2026\-10\-10T18:00:57\.55132\+00:00: rank \#107 2026\-10\-10T19:00:52\.379463\+00:00: rank \#106 2026\-10\-10T20:00:21\.993912\+00:00: rank \#106 2026\-10\-10T21:01:10\.657738\+00:00: rank \#107 2026\-10\-10T22:00:50\.433569\+00:00: rank \#106 2026\-10\-10T23:01:00\.95027\+00:00: rank \#107 ds4 proposes local frontier inference via asymmetric 2\-bit quantization and SSD\-backed KV cache, but a commenter warns the quantized DeepSeek V4 checkpoint is not very good\. ## The brief DwarfStar 4 (ds4) is a C inference engine that aims to run frontier open\-weight models locally on high\-memory Mac, CUDA and ROCm machines\. It supports DeepSeek V4/V4\.1 Flash, GLM 5\.x and Qwen3\.8 Flash Next, using asymmetric 2\-bit quantization for routed experts and an SSD\-backed KV cache that resumes long prefixes by prompt hash\. The project reports benchmark throughput on M5 Max and DGX Spark and exposes a CLI, HTTP APIs and a native coding agent\. The central claim is that compression plus disk\-resident caching makes large MoE models practical on local hardware, though the page is promotional and the benchmarks are self\-reported\. - ds4 is written in C and targets Apple Silicon, NVIDIA CUDA/DGX Spark, and AMD Strix Halo ROCm\. - Supported models include DeepSeek V4/V4\.1 Flash, GLM 5\.x and Qwen3\.8 Flash Next, with text and vision input; supported GGUF layouts only\. - Asymmetric 2\-bit quantization compresses routed experts while keeping critical shared paths precise; imatrix is used\. - The KV cache persists to SSD and reloads by SHA1 of the rendered prompt prefix, avoiding full re\-prefill after restart\. - Reported benchmarks: M5 Max 128 GB q2 at 2,048 tokens: 790\.2 prefill t/s, 39\.4 generation t/s; at 65,536 tokens: 398\.5 prefill, 27\.6 generation\. DGX Spark 128 GB q2 at 2,048: 825\.8 prefill, 18\.1 generation; at 65,536: 823\.0 prefill, 13\.8 generation\. - At 128 GB, GLM 5\.3 Q2 and Qwen Q4 also fit; V4\.1 Q2 streams from SSD\. ## Discussion themes Analyzed: 2026\-10\-03T16:01:40\.887424\+00:00 Analysis sample: Based on 18 of 18 usable stored comments. Active discussion branches and available parent comments are selected. This sample may omit parts of the full thread. Selected themes do not measure community opinion or how common a view is. ### Target hardware and use cases Comments discuss ds4's high\-end hardware targets such as DGX Spark, AMD Ryzen AI Halo and x090 cards, note that less\-equipped users may prefer club\-3090, and request Intel support and improved AMD support\. Sources: [Comment 49939174]() · [Comment 49939079]() · [Comment 49939537]() ### Quantization quality and hardware support gaps One comment says the DSV4 checkpoint makes the quantized model not very good; another notes AMD support needs improvement and Intel support is absent\. Sources: [Comment 49937236]() · [Comment 49939079]() ### Alternative local inference engines and setups Comments mention a DwarfStar\-inspired Intel Xe\-LP engine, club\-3090 for more modest hardware, CPU\-MoE mode with GPU/CPU/SSD offload, and extracting single\-file inference from llama\.cpp\. Sources: [Comment 49938078]() · [Comment 49939174]() · [Comment 49939537]() · [Comment 49939552]() ### Performance and context\-length reports Comments report fused TQ enabling 1M context on a 128 GB M5 Max with Qwen3\.8 Flash Next, and about 22 t/s on an Intel Ultra 7 255H iGPU with a DwarfStar\-inspired engine\. Sources: [Comment 49938274]() · [Comment 49939552]() ### C versus Rust language choice Comments ask why antirez chose C over Rust, explain familiarity and performance goals, and criticize the choice as familiarity\-driven\. Sources: [Comment 49938925]() · [Comment 49939049]() · [Comment 49941410]() ## Sources & coverage AI-generated summary · 2026\-10\-03T08:01:28\.325808\+00:00 Based on 4 of 4 usable stored comments, selected by depth and branch activity. This is a sample of the discussion. Article text may also be shortened. Generated using deepseek\-ai/DeepSeek\-V4\.1\-Flash. Check the linked sources for full context.