{"slug": "a-2014-xeon-vs-two-mac-studios-on-a-125b-model-who-wins", "title": "A 2014 Xeon vs Two Mac Studios on a 125B Model: Who Wins?", "summary": "A €638 HUANANZHI X99-T8 build pairing a 2014 Xeon E5-2673 v3 with 128 GB of DDR3-1600 ECC memory ran the 4-bit Qwen3.8-Flash-Next 125B mixture-of-experts model at 28.9 tokens per second on Strata 0.2.13, matching the 28.6–30.7 tokens per second two networked PCs reached in the author's earlier test. The same machine hit 46 tokens per second on the 3.5-bit Qwen model, 82% of the 56.5 tokens per second from the author's two-RTX 5060 Ti main PC, and 9.3 tokens per second on 2-bit DeepSeek V4.1-Flash. The author reports the Mac Studios were two to almost four times faster but cost four to eight times as much, and credits Claude (Anthropic) with help running the tests.", "body_md": "128 GB of DDR3 on a Chinese X99 board against Apple’s M5 Max and M5 Ultra. The Macs are two to almost four times faster. The X99 costs a quarter to an eighth as much.\n\nWritten with Claude (Anthropic), which also helped me run the tests. Who did what is at the end.\n\nThis is the third part of a series about running very large language models on hardware nobody wants any more. In the first, a 2013 X79 workstation ran Qwen3.8-Flash-Next, a 125-billion-parameter mixture-of-experts model, at 38 tokens per second. In the second, I needed two PCs on gigabit Ethernet to run its 4-bit version at a usable speed.\n\nThis time I moved up one platform generation. The X99 board arrived on 11 November, I built the machine on the 13th, and on the morning of the 14th, with the lab nice and cool, I ran the whole test series on it: three models, the same engine, the same scripts.\n\nIt runs the 4-bit model on its own as fast as the two PCs did together. And while I was reading Mac Studio reviews, I realized I had not seen anyone put a machine like this next to one.\n\nThe machine and what it costs\n\nThe board is a HUANANZHI X99-T8, one of the Chinese boards that run Intel’s 2014 server Xeons with DDR3 memory instead of DDR4. Mine came with a Xeon E5–2673 v3 already fitted: Haswell-EP, 12 cores and 24 threads at 2.4–3.1 GHz. Unlike the X79’s Ivy Bridge, it has AVX2 and FMA. Eight 16 GB DDR3–1600 registered ECC modules fill its four memory channels: 128 GB, with 51 GB/s of bandwidth on paper. The GPUs are an RTX 4070 and an RTX 3060, 12 GB each, both on PCIe 3.0 x16 slots, capped at 170 and 150 W.\n\nThe platform without GPUs costs €638, less than one new RTX 5060 Ti 16 GB: the cheapest one on Amazon Italy costs €800 this week. The two GPUs, which I already owned, are 59% of the total. The 1 TB SSD is enough for all three models below: they take about 450 GB together.\n\nThree models, one morning\n\nAll three runs used Strata 0.2.13, the open-source (MIT) engine I used in the previous articles. It keeps the most-used experts on the GPUs, keeps the rest in RAM where the CPU computes them, and drafts tokens ahead with the model’s own prediction layer (MTP).\n\nThe models:\n\nQwen3.8-Flash-Next (125B parameters, 6B active per token) at 3.5 bits (IQ3_S) and at 4 bits (Unsloth’s UD-Q4_K_XL).\n\nDeepSeek V4.1-Flash, a much bigger mixture of experts (40 layers of 384 experts, six used per token), in Strata’s official 2-bit set. Its files take about 250 GB, including the 111 GB n-gram table.\n\n“Tokens/s” is measured on the whole request, overhead included. “Decode only” is the engine’s own timing while it writes. I use the first one everywhere below, because it is what I wait for.\n\nThe 3.5-bit model runs at 46 tokens per second. My main PC, with two RTX 5060 Ti 16 GB, a Ryzen 7 3700X and 64 GB of DDR4, does 56.5 under the same chat settings (256K context, vision on). The X99 gets 82% of it.\n\nThe 4-bit model is the X99’s real result. Strata wants all 77 GB of its experts in RAM. My main PC has 64 GB, so it reads part of them from the SSD and runs at 16.2 tokens per second. In the previous article, two PCs together reached 28.6–30.7. The X99 alone does 28.9, because 128 GB of DDR3 holds every expert. Here, RAM capacity mattered more than anything else, and it was also the cheapest part of the machine.\n\nDeepSeek runs at 9.3 tokens per second, just under the 10 I consider usable for real work. The day before, I tried it with an experimental branch of Strata that lets a second PC lend its GPU as an expert server over gigabit Ethernet. On that branch the X99 alone did 9.7. With my main PC holding 1,851 experts (15.9 GB) it reached 11.5 tokens per second, and 12.2 when the experts in VRAM were allowed to follow the conversation.\n\nOn code, the 2-bit DeepSeek did at least as well as the two Qwen files. I gave each model twelve programming exercises with unit tests: an expression evaluator, Roman numerals, wildcard matching, date arithmetic and so on. The Qwen models ran them twice, DeepSeek once. DeepSeek failed only the expression evaluator, which every model failed. With twelve exercises, none of these differences is more than noise: not 18 against 19 between the two Qwen files, and not DeepSeek’s 11 out of 12 against them. The failures do differ: the 3.5-bit model got the Gregorian dates wrong and the wildcards right, the 4-bit model the opposite.\n\nPower\n\nI have no wall meter, so these are the components: the CPU package and the RAM from Intel RAPL, the GPUs from the NVIDIA driver.\n\nWith a model loaded and nothing to do, the machine sits at 73–74 W, and at 90 W with DeepSeek. Almost half of the idle power is the RAM: eight registered DDR3 modules draw 33 W doing nothing. Adding the power supply’s losses and about 25 W for the board, disks and fans, I estimate about 320 W at the wall while generating, and about 100 W idle. The CPU never went above 47 °C.\n\nWhat a Mac Studio does with the same models\n\nApple sells the Mac Studio as the machine for local AI, and on these models it is fast. I did not measure the Macs, and none of their numbers comes from the same test as mine: other prompts, other sampling, and on the Ultras another engine. The ratios below are approximate, and I give both ends wherever the choice of figure changes them.\n\nI found three sets of numbers I trust:\n\nM5 Max 128 GB: Strata’s own measurements (MAC.md, DEEPSEEK_MAC.md), with Strata’s Metal engine and the same quantizations I used.\n\nM5 Ultra 256 GB and M3 Ultra 512 GB:MacStories’ review, with the oMLX engine and a 4-bit MLX build of the same Qwen model.\n\nPrices: Italian list prices.\n\nHow the tests differ:\n\nStrata measured the M5 Max on a 200-token story with greedy decoding. I took the story figure; on a short question the same Mac did 101.9 (3.5-bit) and 94.0 (4-bit).\n\nMacStories measured the Ultras at temperature 0 with thinking off. I took the 512-token answer after a 16K prompt; on short prompts the M5 Ultra did about 108 and the M3 Ultra about 70.\n\nThe Mac figures are decode rates. Mine are whole requests: seven different prompts at temperature 0.7 with thinking on. My decode-only rates are 55.5 and 34.2.\n\nThe first two choices flatter the X99 and the third flatters the Macs.\n\nWhat the table says:\n\nThe M5 Max costs 3.8 times as much and is 2 to 2.6 times as fast, or up to 3.3 times at 4 bits on Strata’s short question. At 3.5 bits it reads prompts about as fast as the X99. At 4 bits it reads them 1.5 to 1.7 times faster.\n\nThe M5 Ultra costs 8.1 times as much and is 3 times as fast at 4 bits, or 3.7 times on short prompts. It reads prompts 3.7 times faster.\n\nThe base M5 Ultra (€6,699, 96 GB) cannot keep the 4-bit model in memory. Strata’s Mac page asks for about 91 GB resident and recommends a 128 GB Mac for this file. MacStories’ MLX build peaked at 155 GB of process memory. The 256 GB option adds €4,400, and the configuration MacStories tested, with the 80-core GPU, costs €12,529.\n\nThe M3 Ultra is closer than I expected. It was Apple’s top machine for local AI until September. On the 4-bit model it writes 1.9 times as fast as the X99, or 2.4 times on short prompts, and reads prompts 1.5 times as fast.\n\nDeepSeek is where the gap is smallest. The M5 Max writes 15 to 16 tokens per second at 2 bits, 1.6 to 1.7 times the X99, and reads prompts about twice as fast. This model does not fit in the Mac’s memory: Strata keeps about 68 GB of its experts resident there and reads the rest from the SSD.\n\nPer euro, on the table’s figures, the X99 is 1.5 to 1.9 times better than the M5 Max and 2.7 times better than the M5 Ultra. With the Macs’ short-prompt figures, that narrows to 1.2 to 1.7 times and to 2.2 times.\n\nThe electricity argument\n\nThe usual answer is that the Mac pays for itself on the power bill. It does draw much less: Apple lists 7 W idle for the M5 Max and 9 W for the M5 Ultra, against my ~100 W. So I did the sums.\n\nNothing in this section is measured. The X99’s wall power is my estimate, the Macs’ power while generating is my assumption, and the years below come from a simple scenario, not from a meter. They show the order of magnitude, not a result.\n\nAssumptions:\n\nElectricity price: €0.30 per kWh all-in. That is deliberately high, which favours the Mac.\n\nX99: about 320 W at the wall while generating and 100 W idle, my estimates.\n\nMacs while generating: 90 W for the M5 Max and 200 W for the M5 Ultra, against Apple’s maximums of 200 and 385 W. These are my assumptions.\n\nModel: the 4-bit Qwen, at the speeds in the table.\n\nThe Mac’s lower power draw pays back its price only if the machine generates around the clock for years. For anyone who uses a model the way I do, a few hours a day, it never does.\n\nThis does not depend on my guesses for the Macs. If a Mac used no electricity at all, the X99’s whole bill in the first scenario would be about €225 a year: 19 years to pay back the M5 Max, 49 for the M5 Ultra. Generating nonstop, the X99’s electricity costs €0.93 per million tokens: about 5 years for the M5 Max and 13 for the Ultra.\n\nWhat the Mac does buy is real:\n\ntwo to almost four times the speed, depending on the Mac and the test;\n\nnear silence;\n\na box that fits on a desk;\n\na warranty, new parts, and a resale value in five years;\n\nwith the Ultra, prompts read 3.7 times faster.\n\nIf you run coding agents all day and wait on every answer, that is worth paying for. If you want a 125B model to talk to, the X99 runs the same model at a quarter to half the speed, for an eighth to a quarter of the money.\n\nThe cheap part, again\n\nAt the end of the X79 article I wrote that the cheap part is the one nobody is looking for. The X99 says it again, louder:\n\nRAM: the 128 GB of DDR3 that made the 4-bit model possible cost €100. On a Mac, going from 96 to 256 GB costs €4,400.\n\nBoard and CPU: both together cost €179.\n\nThe expensive parts in this machine are the two GPUs, and they are mid-range cards from 2021 and 2023.\n\nThe machine stays on my bench as a working computer: it already runs a copy of my daily chat, on the 3.5-bit model, with 128 GB of RAM behind it.\n\nMethod\n\nEngine: Strata 0.2.13 with its own packaged engines, int8 KV cache, KV resident window of 32K, vision on, MTP drafting (4 draft tokens; 61–62% accepted on the Qwen models). The CPU’s sustained power limit was set to 105 W; the GPUs were capped at 170 W (RTX 4070) and 150 W (RTX 3060). The 3.5-bit run used the same configuration as my daily chat (256K context); the 4-bit and DeepSeek runs used 128K.\n\nModel files: ISTA-DASLab’s IQ3_S, Unsloth’s UD-Q4_K_XL (the original release that Strata pins; a 111 GB download, 77 GB of it routed experts) and Strata’s official DeepSeek V4.1-Flash 2.00 bpw set, all installed by Strata’s setup and checked against its hashes. Sizes are in decimal GB. The 3.5-bit files were on a SATA SSD; the others on an NVMe drive that this board connects at PCIe 2.0 x4 (1.5 GB/s). Start-up times depend on that.\n\nSpeed: seven different prompts, 400 tokens each, thinking on, temperature 0.7. The first run is discarded to let the expert cache settle, and I report the median of the other six. Ranges: 39.7–50.6 (3.5-bit), 25.3–32.6 (4-bit), 7.8–9.9 (DeepSeek). “Decode only” is the engine’s own timing without request overhead.\n\nLong prompt: a 17,688-token log (16,293 for DeepSeek’s tokenizer) with a password hidden in it, not seen by the engine before.\n\nCoding test: twelve Python exercises with unit tests, temperature 0.2, thinking off, up to 3,000 tokens per answer; two rounds for the Qwen models, one for DeepSeek.\n\nQuality figures are Strata’s, not mine: QUANTS.md for Qwen (top-1 agreement with Q8_0) and DEEPSEEK41.md for DeepSeek (against the original model).\n\nPower: CPU package and DRAM energy from Intel RAPL counters over each phase; GPU power from nvidia-smi every second. Board, disks, fans and power-supply losses are not measured. The wall figures (components ÷ 0.9 + 25 W) are estimates.\n\nMain PC comparison: Ryzen 7 3700X, 64 GB DDR4, 2× RTX 5060 Ti 16 GB, the same Strata files; 3.5-bit at my chat settings, 4-bit measured alone for the previous article.\n\nExpert server test: 13 November, Strata’s experimental dsv41-vision-link branch built from source, my main PC as the expert server over gigabit Ethernet.\n\nMac numbers: not measured by me.\n\nM5 Max, Qwen: Strata’s MAC.md (macOS 27.2, greedy, 200 generated tokens, MTP on, decode rate; engine version not recorded). Story: 94.7 (IQ3_S) and 73.7 (UD-Q4_K_XL). Short question: 101.9 and 94.0. Prompts: the warm rate on 4K and 32K prompts. That page reports contexts up to 32K.\n\nM5 Max, DeepSeek: Strata’s DEEPSEEK_MAC.md (2.00 bpw, engine 0.2.7.1, six experts per token): 15.2–16.2 tokens/s on short answers, 13.8–15.1 after a 14,000-token prompt, which it read at 479–512 tokens/s. Another model’s server was loaded on that Mac during the test, so an idle Mac may be faster.\n\nM5 Ultra 256 GB and M3 Ultra 512 GB: MacStories, oMLX 0.7.0, Qwen3.8-Flash-Next oQ4e, temperature 0, thinking off, multi-token prediction on. Writing speed is a 512-token answer after a 16K prompt; reading speed is on a 15,519-token prompt. Short-prompt prose: about 108 (M5 Ultra) and 70 (M3 Ultra).\n\nMac power: Apple’s published idle and maximum figures, which Apple measured on a 36 GB M5 Max and a 512 GB M5 Ultra; the generating figures are my assumptions.\n\nElectricity: 4-bit speeds (28.9, 73.7 and 87.8 tokens/s). “The same work” means the Mac produces the tokens the X99 produces flat out, then idles. The first scenario gives both machines the same two hours of generation.\n\nPrices: what I paid for the board with the CPU, the RAM and the power supply; the GPU, SSD, cooler and case prices are my estimates of today’s market. The RTX 5060 Ti 16 GB price is the cheapest new card on Amazon Italy in mid-November 2026. Mac prices are Italian Apple list prices at launch: M5 Max with 128 GB and a 512 GB SSD €5,909 (€6,239 with 1 TB); M5 Ultra 96 GB €6,699; M5 Ultra with 36-core CPU, 80-core GPU, 256 GB and 1 TB €12,529.\n\nHow this was made\n\nI built the machine and decided what to measure. Claude (Anthropic) worked with me on the tests and on the text. On the testing side, it helped me set up and run the test series on my machines and read the logs. On the editing side, it wrote the article from my notes and measurements, checked every number taken from Strata’s documentation, MacStories and Apple against its source, and recomputed the ratios and the electricity sums. I reviewed and edited the text. Every measurement of the X99 and of my main PC was run on my own machines.", "url": "https://wpnews.pro/news/a-2014-xeon-vs-two-mac-studios-on-a-125b-model-who-wins", "canonical_source": "https://pub.towardsai.net/a-2014-xeon-vs-two-mac-studios-on-a-125b-model-who-wins-74aabdfa603a?source=rss----98111c9905da---4", "published_at": "2026-11-15 16:31:01+00:00", "updated_at": "2026-11-15 16:51:48.284362+00:00", "lang": "en", "topics": ["large-language-models", "ai-infrastructure", "ai-chips"], "entities": ["HUANANZHI", "Intel", "Xeon E5-2673 v3", "Apple", "Mac Studio", "Qwen3.8-Flash-Next", "DeepSeek V4.1-Flash", "Strata"], "also_reported_by": [], "alternates": {"html": "https://wpnews.pro/news/a-2014-xeon-vs-two-mac-studios-on-a-125b-model-who-wins", "markdown": "https://wpnews.pro/news/a-2014-xeon-vs-two-mac-studios-on-a-125b-model-who-wins.md", "text": "https://wpnews.pro/news/a-2014-xeon-vs-two-mac-studios-on-a-125b-model-who-wins.txt", "jsonld": "https://wpnews.pro/news/a-2014-xeon-vs-two-mac-studios-on-a-125b-model-who-wins.jsonld"}}