Nvidia, You’re Late. World’s First 128GB LLM Mini Is Here!

Nvidia's DJX Spark is still weeks away, always 'coming soon' — Opinion (50/100)

At 0:00

Sassy shade on the eternal 'coming soon' — fair roast on delays 😏

Why this score: This is commentary on Nvidia's release timeline, not a hard fact claim. Speaker mocks the repeated delays, which aligns with real-world patterns for tech launches. *No specific date given to verify*, but the sarcasm lands as opinion.

Original quote: “[0:00] DJX Spark will be ready, will be [0:02] available shortly, probably in a few [0:04] weeks. DJX Spark is coming. It's always [0:07] been coming. But while we wait, GMK Tech [0:10] is not sitting around. They just dropped [0:12] the Evo X2”

GMKtec Evo X2 has 128GB unified RAM for LLMs, half price of DJX Spark — Solid (85/100)

At 0:12

👌💯

Why this score: Claims match product specs and market positioning. - GMKtec EVO-X2 confirmed with AMD Ryzen AI Max+ 395, 128GB LPDDR5X unified memory for local LLMs. - Positioned as cheaper alternative to Nvidia DGX Spark (note: likely 'DGX', transcript typo on 'DJX'). *Price comparison directionally accurate* per 2025 launch coverage.

Original quote: “[0:12] the Evo X2, a mini PC with 128 GB of RAM [0:18] that's shared with the GPU, ready for [0:20] local LLMs and machine learning, and it [0:23] costs less than half the price of the [0:25] DJX Spark. Hey, maybe the Spark is out [0:27] by the time this video is edited. And in [0:29] that case,…”

Dual PCIe Gen4 SSDs up to 16TB, Wi-Fi 7, shareable memory via static partitioning — Solid (75/100)

At 0:30

Checks out 👌

Why this score: Hardware details accurate. - Dual PCIe Gen4 SSD slots (up to 2TB noted, 16TB possible with multiples); Wi-Fi 7 standard. - Correctly notes non-true unified memory, uses static partitioning/sharing unlike Apple Silicon.

Original quote: “It also has dual PCIe Gen 4 SSD slots inside. So, yeah upgradeable, too, up to 16 TB. And it's got Wi-Fi 7 ... Well, it's not exactly unified memory. The memory can be shared, but it's different from Apple Silicon's or unified memory architecture. It's called static partitioning.”

Testing Asus ROG Flow Z13 with 128GB chip — Solid (80/100)

At 2:30

👌

Why this score: Asus ROG Flow Z13 exists with 128GB configs. - 2025 model uses AMD Ryzen AI Max+ 395, up to 128GB LPDDR5X RAM per web context. - Matches video's '128 gigs on this device' demo.

Original quote: “[2:28] this device. This is Asus's Flow Z13 [2:31] but the thermals on these machines are [2:33] very, very different. So, I wonder if”

LM Studio offloads all 64/64 layers of Q4 Quinn 32B to GPU — Solid (75/100)

At 3:00

✅

Why this score: LM Studio layer offloading is standard. - App supports GPU offload for quantized LLMs like 32B Q4 (~18GB). - 128GB unified setup enables full offload, as shown in task manager.

Original quote: “[3:03] quantization is four. And LM Studio is [3:06] telling me I can offload 64 out of 64 [3:10] layers. So, let's do that. Load model.”

Static partitioning: copies from system to GPU memory, unlike Apple's unified — Verified (90/100)

At 3:45

Spot on 👌

Why this score: Accurate distinction in memory architectures. - Discrete GPUs use static partitioning (separate VRAM/RAM, data copied via PCIe). - Apple Silicon has unified memory (shared pool for CPU/GPU). Matches demo and web context.

Original quote: “[3:25] that's what this static partitioning of [3:26] memory is all about. It's not unified [3:28] memory like it is on Apple. We'll see”

Quinn 32B Q4 is 18.4GB, gets 10.58 tokens/sec — OK (65/100)

At 4:20

Plausible benchmark ⚠️

Why this score: Reasonable perf for hardware. - 32B Q4 typically ~18-20GB; 10.58 t/s fits Ryzen AI Max+ 395 with Radeon iGPU. - No exact 'Quinn' model match, but generic 32B Q4 stats align; live demo credible.

Original quote: “[4:00] tokens per second here, which is pretty [4:02] reasonable for a 32 billion parameter [4:04] model that's 18.4 GB in size. Let's do”

Offloading LLM layers to CPU makes it very slow — Solid (85/100)

At 4:30

👌✅

Why this score: Standard LLM best practice confirmed. - GPUs dominate for LLM inference due to parallel compute needs; CPUs slower for transformers. - Web context verifies: GPUs essential for high performance in large models.

Original quote: “the way, if you're kind of new to this if you offload even any layers to the CPU, things are just going to go downhill really fast and become very very slow. You want everything to be on GPU.”

Apple M4 chip has 120 GB/s memory bandwidth — Verified (95/100)

At 4:40

Spot on number ✓

Why this score: Apple's official spec matches exactly. - Base M4 Mac Mini (2024) advertised at 120 GB/s; measurement of 96 GB/s triad in STREAM is realistic (not peak). - Higher in Pro/Max variants, but base confirmed.

Original quote: “according to Apple, the M4 chip has the memory bandwidth of 120 GB per second. Pretty good already. And that's just the base model.”

Measured 96 GB/s triad on M4 STREAM benchmark — Personal Story (70/100)

At 5:00

Plausible real-world measurement ⚠️

Why this score: Speaker's hands-on test aligns with expectations. - STREAM triad often ~80% of peak bandwidth; 96 vs. 120 GB/s reasonable for sustained load. - STREAM benchmark industry standard, referenced by AMD.

Original quote: “So, right now, I measured this running stream. And if we look at this Triad number, we're getting 96 GB per second or thereabout. It's not 120 like Apple advertises, but it's pretty close.”

DeepSeek R1-Distill-Qwen-7B Q4 is ~4 GB — Solid (80/100)

At 5:25

👌

Why this score: Model details accurate. - DeepSeek R1-Distill-Qwen-7B is real 7B param model; Q4 quantization typically ~4-5 GB for 7B models. - Matches web context on size and use.

Original quote: “I've got DeepSeek R1 distill quen 7 billion parameter model with a Q4 quantization. It's about 4 GB on disk.”

M4 Mac Mini at 31W, 19.94 tokens/sec — OK (60/100)

At 6:30

Specific benchmark shown — no contradiction but no public match either ⚠️

Why this score: Plausible demo numbers. - On-screen metrics match what's said for M4 Mac Mini under LLM load. - Web benchmarks vary by model/workload; no exact public replica found, but consistent with Apple M4 silicon efficiency.

Original quote: “And I'm going to have it write a little bit of a longer output. And we are at 31 watts, which is actually pretty decent. What am I saying? Decent. I mean, this is pretty good, right? 31 watts. We got 19.94 tokens per second.”

Llama CPP compiles for ROCm and Vulkan — Solid (85/100)

At 8:30

✅

Why this score: Accurate description of Llama.cpp compilation options. - Llama.cpp supports multiple backends including ROCm (AMD), Vulkan (cross-platform), CUDA, and CPU-only. - LM Studio uses Llama.cpp as its core inference engine with these runtime options.

Original quote: “[8:30] on what LM Studio runs on. And Llama CPP [8:33] when compiled for the platform, you can [8:35] compile it uh for Rockom, you can”

ROCm not available yet for this setup — Dubious (45/100)

At 9:15

ROCm 'not available'? Tell that to AMD's 2026 docs 🚩

Why this score: Outdated claim about ROCm availability. - ROCm has supported Llama.cpp since at least early 2026 with 'Day One' support for Gemma 4 models. - Available on Linux (Ubuntu 24.04/22.04) for AMD 9000 series GPUs and Instinct hardware. - Speaker's 'coming soon' status doesn't match current April 2026 reality.

Original quote: “[9:15] something called Rockom for AMD [9:18] processors. Unfortunately, Rockom is not [9:21] available and I'll come back to that.”

Sources: Compatibility matrix - ROCm Documentation - AMD, Limitations and recommended settings — Use ROCm on Radeon GPUs, Installation prerequisites — ROCm installation (Linux) - ROCm Documentation - AMD

Vulkan Llama.cpp: 935 PP, 46 TG tokens/sec — OK (65/100)

At 10:00

Plausible benchmark numbers for Vulkan on beefy AMD hardware ⚠️

Why this score: Benchmark results are realistic for the hardware. - 935 tokens/sec PP and 46 TG/sec align with Vulkan Llama.cpp performance on high-end AMD GPUs (RX 7900+). - Matches LM Studio's reported 43 TG/sec within margin of error as speaker notes. - Specific model (DeepSeek R1 Q4) and hardware context makes these credible.

Original quote: “[10:07] 935 tokens per second. [10:10] That's nice. And we're getting TG128. TG [10:13] is token generation and that's 46 tokens [10:16] per second. So, we're kind of close to”

M4 Mac Mini gets 228 PP512, 21 tokens/sec on Llama Bench — Personal Story (70/100)

At 10:30

Dude's own benchmarks — fair play 👌

Why this score: Speaker shares personal test results from running Llama Bench on their M4 Mac Mini with llama.cpp. - Llama Bench is a real tool for LLM inference benchmarking, as confirmed by project docs. - Anecdotal but consistent with tool's use on Apple Metal backend; no contradiction in web context.

Original quote: “Running Llama Bench with the same model on the M4 Mac Mini, we get a much much lower result. 228 for PP512 and TG128 gets us 21 tokens per second, which uh kind of aligns with what we've seen in LM Studio on the Mac as well.”

M4: 120GB/s, M4 Pro: 273GB/s, M4 Max: 540GB/s, M3 Ultra: 819GB/s memory bandwidth — Verified (100/100)

At 11:36

✅

Why this score: Exact match to official specs. - M4: 120 GB/s confirmed. - M4 Pro: 273 GB/s confirmed. - M4 Max: 546 GB/s (speaker's '540 or around there' is precise enough). - M3 Ultra: 819 GB/s confirmed.

Original quote: “For the M4, it's 120 GB per second. For the M4 Pro, it's 273 GB per second. For the M4 Max, it's 540 GB per second or somewhere around there. The M3 Ultra 819 GB per second.”

Nvidia DGX Spark has 279 GB/s memory bandwidth — Solid (80/100)

At 11:54

👌

Why this score: Plausible and aligns with video overview (DGX Spark from CES 2025). - Speaker cites 279 GB/s, slightly above M4 Pro (273 GB/s). - Web context confirms Nvidia's DGX Spark as competitor in local LLM space; specific bandwidth not contradicted.

Original quote: “And where's Nvidia's DJX Spark on that? Well, that's at 279.”

This machine's bandwidth at 279 GB/s, higher than M4 Pro — Solid (80/100)

At 12:30

👌

Why this score: Matches official specs. - M4 Pro official bandwidth: 273 GB/s; transcript's 279 is close, likely measured. - Web context confirms M4 Pro at 273 GB/s.

Original quote: “Well, that's at 279. So, slightly higher memory bandwidth than the M4 Pro.”

M4: 96/120 GB/s Stream; M4 Pro: 209/273 GB/s — Solid (85/100)

At 12:48

✅

Why this score: Real benchmark results align with expectations. - Stream often shows < theoretical max (e.g., M4 96/120, M4 Pro 209/273). - Web context notes typical variance in Stream for Apple/AMD chips.

Original quote: “Over here on the M4, I got 96 GB out of the 120 that Apple advertises. ... M4 Pro, not bad. 209 out of the 273 that Apple says.”

This machine's Stream benchmark: 120 GB/s vs expected 256 — OK (65/100)

At 13:15

Low real-world vs spec — common but underwhelming ⚠️

Why this score: Plausible benchmark result. - AMD Ryzen AI Max (e.g., Framework Desktop) specs ~256 GB/s theoretical LPDDR5X. - Stream often 40-50% of max due to config/benchmark limits; 120 GB/s fits pattern. - Speaker's own test, transparent about discrepancy.

Original quote: “What about this machine? ... And we're getting 120. 120 is way way lower than I would expect because we wanted 256, right?”

Gemma3-1B: ~160 t/s; Llama3.2-1B: 130 t/s on this machine — OK (60/100)

At 13:51

Self-reported benchmarks — impressive if real 📈

Why this score: Speaker's hands-on Llama Bench results via Vulkan. - High tokens/sec for small models align with AMD Ryzen AI Max+128GB unified mem claims. - No public matching data yet (new hardware), but consistent with video thesis on local LLM power. - *Unverified externally but transparently presented.*

Original quote: “Gemma 3 1 billion. We're getting 100 almost 160 tokens per second here. Llama 3.2 1 billion. Instruct 130.”

Gemma 3 12B: 23.9 t/s, Qwen 2.5 32B: 10.8 t/s, Llama 3 70B Q4: 5 t/s — OK (65/100)

At 14:30

Benchmark numbers look real but Llama 3 '370B' is sus — it's 70B 😂⚠️

Why this score: Speaker reports specific tokens-per-second benchmarks from on-screen tests for real models like Gemma 3 12B, Qwen 2.5 32B, and Llama 3 70B (likely typo for 370B). - Models exist and match descriptions (Gemma 3 multimodal 12B from Google Mar 2025, Qwen2.5-32B Sept 2024, Llama 3 70B April 2024). - Scores reasonable for mini PC hardware; Llama 3 70B at 5 t/s aligns with thermal limits on compact devices, though '370B' phrasing is odd but contextually 70B quantized.

Original quote: “We've seen 44. Gemma 312 billion. Now we're getting into really more useful models. 23.9 still a really decent score. 14 billion 23.7. Quen 2.5 quarter 32 billion. 32 billion parameter models are actually pretty useful. We're getting 10.8. Not the best. I'd rather be over 20, but still 10.8 is…”

Knuckbox Evo X2 desktop beats Flow Z13 laptop in most LLM benchmarks — Dubious (45/100)

At 15:10

'Knuckbox' or 'Nugbox'? And no hits on that exact desktop model 🚩🤔

Why this score: Benchmark comparison references on-screen charts showing green dots (desktop) slightly outperforming orange (laptop) except Gemma 3 12B. - Flow Z13 matches ASUS ROG Flow Z13 (2025) with AMD Ryzen AI Max+ 395 and 128GB LPDDR5X. - 'Knuckbox/Nugbox Evo X2' likely GMKtec EVO-X2 (video thesis), but exact name garbled in transcript and no direct web confirmation of 'Knuckbox' — plausible but unverified hardware.

Original quote: “This is the results for the Knuckbox Evo X2 which is this, versus that FlowZ13 with 128 gigs. And you can see the little green dots are just a little bit higher than the orange dots for each one of these, except for this one. Gemma 312 billion. The laptop beats the desktop for some reason on that…”

M4 Max: 49 t/s Gemma 3 12B; Nvidia Spark Blackwell has 278 bandwidth — Solid (80/100)

At 15:43

👌 Numbers align with unified memory advantages

Why this score: Cross-hardware benchmarks for Gemma 3 12B (Nugbox 23 t/s vs M4 Max 49 t/s) explained by memory bandwidth differences. - Nvidia DGX Spark (Blackwell-based) confirmed with higher bandwidth (~278) vs competitors; partners like ASUS/Dell shipping mid-2025. - Matches video thesis and real 2025 hardware launches.

Original quote: “I have a slightly different view here that compares a bunch of different machines. The FL Z13 with 32 gigs. We've got the Nugbox Evo here. This is Gemma 3 12 billion parameter model, 23 tokens per second. And here's the M4 Max machine with 49 tokens per second for that same model. So quite a bit…”

Static memory partitioning is stable but wastes memory and lacks flexibility for modern PCs — Solid (85/100)

At 16:33

✅

Why this score: Accurate description of static partitioning pros/cons. - Matches standard OS concepts: fixed blocks enhance stability/isolation but cause fragmentation and limit flexibility. - Aligns with web context on static vs dynamic partitioning drawbacks for modern systems.

Original quote: “here we do have static partitioning which is fixed and of course that's a way of dividing memory into set blocks before the computer starts up. So you kind of have to know ahead of time. What's nice about that is easier to set up, uses few system resources and keeps each program in its own space…”

AMD at Computex promised better ROCm support for LLM chips like this one — Solid (80/100)

At 17:29

👌

Why this score: Plausible and aligns with AMD's ROCm expansion efforts. - Web context notes AMD's Computex 2025 announcements on LLM focus; CES 2026 extended ROCm to Ryzen AI Max+. - Speaker's claim fits timeline of AMD prioritizing ROCm for consumer AI hardware.

Original quote: “According to AMD, this is going to significantly improve in theory. ... However, during Computex, they did say they would pay more attention to the LLM community. Hey, that's us. and get Rockom support in more chips, including this one”

Michael Larabel got ROCm working on this chip but unstable with seg faults — OK (65/100)

At 17:59

Specific tester named but sounds like fragile hack job ⚠️

Why this score: Refers to real Phoronix writer Michael Larabel (likely 'Phoronix' misheard). - Matches pattern of early ROCm adopter reports on AMD AI chips: works with specific old versions (ROCm 6.4.1, Ubuntu 24.04, kernel 6.11) but unstable (Blender HIP crashes, Llama.cpp segfaults). - Credible as anecdotal tech reporting, but unverified exact article; common in AMD ROCm ecosystem pre-full support.

Original quote: “You'll see a lot of people trying it on this chip and with limited success. For example, Michael Larabel from Feronx got this working as long as you use Rockom 6.4.1 and install Ubuntu 24.04, 04 not 25 sticking with the Linux 6.11 HWE kernel and avoid trying to build DKMS modules on newer kernels…”

64GB version $1499, 128GB version $2000 — Solid (80/100)

At 18:30

👌

Why this score: Pricing aligns with high-end mini PC market. - GMKtec EVO-X2 (AMD Ryzen AI Max Plus 395, 128GB) positioned at premium ~$2000 due to demand and specs. - Video overview and 2026 trends confirm rising costs for DDR5/128GB configs amid shortages.

Original quote: “$14.99 for the 64 GB version and 2 grand for the 128 GB version that I have here.”

This mini outperforms M4 Mac Mini due to bandwidth, cores, RAM; M4 Pro 64GB costs $2399 — OK (65/100)

At 19:19

Plausible edge on RAM/bandwidth, M4 Pro price ~spot on ⚠️

Why this score: Performance comparison reasonable but unbenchmarked here. - M4 Mac Mini/Pro maxes at lower unified RAM (64GB on Pro); this has 128GB LPDDR5X. - M4 Pro starts ~$1399 but configured 64GB likely nears $2400; bandwidth/cores favor AMD for LLMs per thesis. - Needs direct benchmarks for full verify.

Original quote: “if you compare it to the M4 Mac Mini, you saw that the performance of the Mac Mini is good, but because of the memory bandwidth, limitations,, it's, not, going to be as powerful as this thing. Also, the number of cores is much less, so it's going to be less powerful. Also, 64 and 128 gigs of RAM.…”