[{"data":1,"prerenderedAt":493},["ShallowReactive",2],{"project-50127979":3},{"id":4,"name":5,"line":6,"line_source":7,"title":8,"url":9,"domain":10,"maker":11,"posted_at":12,"year":13,"points":14,"comments":15,"rank_in_year":16,"shown_in_year":17,"top":18,"tier":19,"category":20,"subcategory":21,"format":22,"keywords":23,"reception":28,"sentiment":29,"stance_mix":30,"scored_comments":35,"aspects":36,"strip":45,"post_text":46,"what":47,"said":69,"maker_in_thread":197,"shape":206,"grid":210,"attempts":455,"same_week":456,"maker_shows":42,"maker_hidden":485,"reader_version":486,"card":487},"50127979","Llmspeed","feel how fast a local LLM runs before you buy the hardware","title","Show HN: Llmspeed – feel how fast a local LLM runs before you buy the hardware","https:\u002F\u002Fllmspeed.dev\u002F","llmspeed.dev","albinstman",1792211813,2026,331,86,113,4366,3,"legend","ai","local llm benchmarking","web_app",[24,25,26,27],"local llm","gpu benchmark","tokens per second","llm hardware","curious",-0.198,{"positive":31,"neutral":32,"mixed":33,"critical":34},5,28,2,15,50,{"performance":37,"execution":39,"idea":41,"pricing":43,"design":44},{"positive":18,"critical":38},11,{"positive":40,"critical":33},0,{"positive":42,"critical":42},1,{"positive":40,"critical":40},{"positive":40,"critical":40},[42,40,18,42,18,40,40,31,40,40,42,42,42,31,42,31,40,18,18,18,18,18,18,42,40,40,18,42,18,18,42,18,42,18,18,18,18,42,42,18],null,{"problem":48,"audience":54,"pricing":46,"motivation":62,"time_to_build":46,"stack":66,"openness":67,"stage":68,"team_size":67},{"value":49,"source":50,"evidence":51},"Hard to know if new hardware will run local models at useful speed before buying it","maker",[52],{"id":53},"post",{"value":55,"source":56,"evidence":57},"People considering hardware for running local LLMs","inferred",[58],{"id":59,"by":60,"text":61},"50130918","tom_alexander","This is great. It solves two of my issues: ``` 1. It is difficult to figure out if dropping $10k+ on hardware will make a meaningful difference for running local models. 2.…",{"value":63,"source":50,"evidence":64},"Let people feel how fast a local LLM runs before buying hardware",[65],{"id":53},[],"unknown","launched",{"praise":70,"objections":91,"requests":127,"questions":160,"alternatives":178},[71,80],{"value":72,"source":73,"evidence":74},"Useful for judging whether expensive hardware makes a difference","commenter",[75,76],{"id":59,"by":60,"text":61},{"id":77,"by":78,"text":79},"50128532","pil0u","I don't know what basis these numbers are estimated on, but my M2 Ultra 128GB runs almost twice as fast as what's reported for the M3 Ultra 512GB. I wished there was such a tool…",{"value":81,"source":73,"evidence":82},"Good idea and well designed",[83,87],{"id":84,"by":85,"text":86},"50140351","riobits","It would be cool if I could just search for a GPU instead of scrolling through the whole list. Overall, great website!",{"id":88,"by":89,"text":90},"50129032","kinduff","Cool examples, this is interesting to test. I've noticed that OpenAI and Meta play a lot with animations and fades to mask the delays.",[92,108,116],{"value":93,"source":73,"evidence":94},"Speed estimates are off compared with real setups",[95,99,103,107],{"id":96,"by":97,"text":98},"50134730","kgeist","It says 46 tok\u002Fsec for Qwen3.6-27b 4 bit on RTX5090, but I get 60-70 tok\u002Fsec on a small context like that (5 bit weights with 8 bit KV cache), although it does slow down to 40…",{"id":100,"by":101,"text":102},"50130953","adrian17","A random number I checked feels kinda optimistic, or there's something wrong with my setup; Gemma 26B-A4B on a 16GB 5060Ti is reported to get 21-88tok\u002Fs with Q4_K_M, meanwhile I…",{"id":104,"by":105,"text":106},"50133214","lisplist","Just chiming in that the idea is cool, but the speeds are way wrong",{"id":77,"by":78,"text":79},{"value":109,"source":73,"evidence":110},"Ignores llama.cpp tuning, speculative decoding and context size",[111,112],{"id":96,"by":97,"text":98},{"id":113,"by":114,"text":115},"50129157","mirekrusin","It would be great if numbers weren't off, would include optimised versions (ie. speculative decoding), context size, include allowed parallel inference (single agents use…",{"value":117,"source":73,"evidence":118},"Model and quant choices are limited and default to tiny quants",[119,123],{"id":120,"by":121,"text":122},"50132508","layla5alive","This tool has very tiny quants available for each model, and defaults to the tiny quants, so I'm not so sure you... tried it?",{"id":124,"by":125,"text":126},"50128529","bozdemir","Model list is a bit limited, and whole harness affects the speed, so the libs used in harness and their versions might be useful to show for each individual model setup.",[128,143,149],{"value":129,"source":73,"evidence":130},"Multi-GPU configurations",[131,135,139],{"id":132,"by":133,"text":134},"50128915","gitowiec","I would like to see in such simulators: - multi GPU configuration (hetero and homo) - PCIe gen chooser (3.0\u002F4.0\u002F5.0) - is nvlink enabled (2 cards, 2 x 2 cards, 3 x 2 cards) - is…",{"id":136,"by":137,"text":138},"50134592","sleight42","Please modify to include multiples of each as well. Most serious people aren't going to use a single PCIe card if they can help it.",{"id":140,"by":141,"text":142},"50128804","Perenti","Doesn't provide for hybrid setups RTX5070 + RTX5080 for example.",{"value":144,"source":50,"evidence":145},"Add harness details to results",[146],{"id":147,"by":11,"text":148},"50128660","I’ll look into adding harnesses. The benchmark sources usually don’t contain harness so the problem is limited data. What models would you like to see?",{"value":150,"source":73,"evidence":151},"More models and hardware such as DGX Spark laptops",[152,156],{"id":153,"by":154,"text":155},"50131201","ilaksh","A ton of RTX Spark (supposedly similar to DGX Spark) laptops just came out. We really need this for those. Asus ProArt P14, Surface Ultra, Yoga Pro 9n, etc. I am hoping the…",{"id":157,"by":158,"text":159},"50130737","Tepix","Missing: Dual Strix Halo, Dual Nvidia Spark",[161,167],{"value":162,"source":73,"evidence":163},"How accurate are the simulations?",[164],{"id":165,"by":166,"text":162},"50138890","prantosmss",{"value":168,"source":73,"evidence":169},"How does it handle context length and prefill?",[170,174],{"id":171,"by":172,"text":173},"50134555","Geee","How about context length? Isn't it still just a few thousand tokens on local models?",{"id":175,"by":176,"text":177},"50128720","friendly_chap","This is a great idea but I'm far less concerned about output token speed and way more about the ability to parallelize prefill - for agentic coding that's pretty much way more…",[179,186],{"value":180,"source":73,"evidence":181},"Sparkbench for platform-specific leaderboards",[182],{"id":183,"by":184,"text":185},"50137941","verdverm","There are also platform specific sites which have leaderboards for different models and setups. Having an OEM Spark, this one has been very helpful https:\u002F\u002Fsparkbench.dev\u002F",{"value":187,"source":73,"evidence":188},"Renting GPUs on vast.ai to test hardware",[189,193],{"id":190,"by":191,"text":192},"50128971","kristopolous","just go rent one for a few minutes and try it out ... it'll cost you well under $0.50 if you're just trying to get a feel for if it's right for you. There's a lot of things like…",{"id":194,"by":195,"text":196},"50134796","montroser","Depending on what hardware you're considering, you might rent a vast.ai box with similar specs for a month or two to get a feel for it.",{"comments":31,"answered":40,"key":198},[199,202,203],{"id":200,"text":201},"50132763","Would love to have multi-gpu support eventually but it's a balance between whats popular and what actually has good quality benchmarks available. Multi-gpu numbers can also vary…",{"id":147,"text":148},{"id":204,"text":205},"50128990","Qwen 3.8 flash next is already in the works and should be live soon.",{"commenters":207,"max_depth":31,"maker_replies":31,"first_comment_mins":208,"lifespan_hours":209},64,118,33,[211,215,219,222,224,228,230,232,235,239,243,246,249,251,253,255,258,262,266,269,272,274,277,279,282,285,289,290,293,298,300,301,305,308,312,315,317,320,323,325,327,330,333,336,338,341,344,346,349,351,354,357,358,361,363,366,368,369,371,375,379,382,384,388,390,394,396,399,401,406,409,413,415,417,419,423,426,428,430,432,435,440,442,446,448,452],[212,40,42,213,214,40],"50128999","project\u002Fcritical\u002Fperformance","lopotubez",[216,40,40,217,218,40],"50129192","commenter\u002Fneutral\u002Fnone","samuell",[220,40,40,217,221,40],"50129393","Abishek_Muthian",[223,40,40,217,218,40],"50129849",[225,40,18,226,227,40],"50129389","project\u002Fneutral\u002Fnone","testerius",[229,40,40,217,214,40],"50129803",[231,40,40,217,227,40],"50129957",[233,40,42,213,234,40],"50140877","zardo",[236,40,40,237,238,40],"50132803","other\u002Fneutral\u002Fnone","oniony",[240,40,18,241,242,40],"50129445","project\u002Fneutral\u002Fexecution","OtomotO",[244,40,40,217,245,40],"50129769","trucks-refinish",[247,40,40,217,248,40],"50129807","PaulKeeble",[250,40,40,217,248,40],"50129784",[252,40,40,217,214,40],"50129914",[254,40,40,217,195,40],"50134733",[256,40,31,257,248,40],"50129753","project\u002Fpositive\u002Fperformance",[259,40,18,260,261,40],"50131291","project\u002Fneutral\u002Fperformance","wilkystyle",[263,40,40,264,265,40],"50131334","commenter\u002Fneutral\u002Fperformance","lelandbatey",[267,40,40,237,268,40],"50139529","anon373839",[270,40,40,217,271,40],"50130913","rhdunn",[273,40,40,217,214,40],"50132086",[275,40,40,217,276,40],"50132126","AntiRush",[120,40,42,278,121,40],"maker\u002Fcritical\u002Fexecution",[280,40,42,213,281,40],"50132971","fennecbutt",[283,40,42,213,284,40],"50138394","bitexploder",[286,42,40,287,288,40],"50131687","other\u002Fneutral\u002Fperformance","Jabdoa2",[96,33,42,213,97,40],[291,18,40,292,195,40],"50134971","other\u002Fneutral\u002Fpricing",[294,295,31,296,297,40],"50128820",4,"project\u002Fpositive\u002Fidea","hizlikovboy27",[299,295,18,226,154,40],"50131211",[175,31,42,213,176,40],[302,303,40,237,304,40],"50136463",6,"Grombobulous",[190,306,18,307,191,40],7,"project\u002Fneutral\u002Fpricing",[309,310,31,257,311,40],"50128569",8,"hbogert",[313,310,40,217,314,40],"50128845","internet_points",[316,310,40,217,311,40],"50129051",[318,310,18,260,319,40],"50133726","merbanan",[321,310,40,217,322,40],"50128849","dotancohen",[324,310,18,260,311,40],"50129043",[326,310,40,217,322,40],"50130072",[132,328,18,329,133,40],9,"project\u002Fneutral\u002Fidea",[331,328,40,217,332,40],"50130724","mkesper",[334,328,18,226,335,40],"50132565","tempoponet",[200,328,40,337,11,42],"maker\u002Fneutral\u002Fnone",[59,339,18,340,60,40],10,"project\u002Fmixed\u002Fexecution",[342,38,31,257,343,40],"50130708","thadk",[113,345,42,213,114,40],12,[183,347,31,348,184,40],13,"project\u002Fpositive\u002Fnone",[100,350,42,213,101,40],14,[352,34,18,260,353,40],"50140138","sgt",[124,355,42,356,125,40],16,"project\u002Fcritical\u002Fnone",[147,355,40,337,11,42],[359,355,40,217,360,40],"50128740","cyanydeez",[362,355,40,217,125,40],"50128901",[364,355,40,217,365,40],"50128917","jimnotgym",[367,355,18,226,125,40],"50128907",[204,355,40,337,11,42],[370,355,40,337,11,42],"50132337",[372,373,40,237,374,40],"50140699",17,"voxleone",[376,377,42,213,378,40],"50131514",18,"binary132",[380,377,18,260,381,40],"50132009","snerbles",[88,383,40,237,89,40],19,[385,386,18,260,387,40],"50136321",20,"layoric",[389,386,18,260,284,40],"50138412",[391,392,18,329,393,40],"50129353",21,"pinkpall",[77,395,42,213,78,40],22,[397,395,40,398,11,42],"50128636","maker\u002Fneutral\u002Fperformance",[153,400,18,226,154,40],23,[402,403,42,404,405,40],"50129556",24,"project\u002Fcritical\u002Fexecution","AYBABTME",[407,403,40,217,408,40],"50129743","numpad0",[410,411,18,329,412,40],"50132286",25,"layer8",[414,411,40,217,281,40],"50134381",[194,416,18,226,195,40],26,[136,418,18,329,137,40],27,[420,32,18,421,422,40],"50129785","project\u002Fneutral\u002Fdesign","arcfour",[84,424,18,425,85,40],29,"project\u002Fmixed\u002Fdesign",[157,427,18,226,158,40],30,[165,429,18,260,166,40],31,[171,431,18,329,172,40],32,[433,209,18,226,434,40],"50131956","hsnewman",[436,437,42,438,439,40],"50134963",34,"project\u002Fcritical\u002Fidea","BatchJob",[104,441,42,213,105,40],35,[443,444,18,226,445,40],"50131733",36,"Dwedit",[140,447,18,329,141,40],37,[449,450,18,421,451,40],"50130351",38,"amelius",[453,454,18,329,451,40],"50130270",39,[],[457,462,467,471,476,481],{"id":458,"name":459,"points":460,"category":461},"50170593","Permit Pending",941,"games",{"id":463,"name":464,"points":465,"category":466},"50102247","Unfenced",319,"web_saas",{"id":468,"name":469,"points":470,"category":20},"50118688","StarSkirmish Bench",254,{"id":472,"name":473,"points":474,"category":475},"50137950","Sphinx",253,"dev_tools",{"id":477,"name":478,"points":479,"category":480},"50140769","The Zelda timeline",231,"learning_science",{"id":482,"name":483,"points":484,"category":475},"50131189","Esp32sim",212,false,"v2",{"id":488,"name":5,"line":6,"ai":485,"cat":20,"year":13,"points":14,"comments":15,"rec":28,"s":489,"top":18,"kw":490,"maker":11,"domain":10,"tier":19,"attempts":40,"strip":491,"url":9,"mix":492,"praise":72,"objection":93,"sub":21},50127979,-0.2,[24,25,26,27],"1031300500111515033333310031331313333113",[31,32,33,34],1794036712291]