artificialintelligence

The NVIDIA DGX Spark is palm-sized, and the system reports 121.93 GiB of usable memory. That number invites a tempting conclusion: if all 128 GB can serve as GPU memory, surely it can run any large model you throw at it?
Fitting a model in memory doesn't mean it will run fast. What this machine can actually do comes down to two numbers, and they measure two very different things.
Two numbers: capacity and bandwidth
The DGX Spark is built on NVIDIA's GB10 chip. The CPU and GPU share a single pool of 128 GB of LPDDR5x memory, and after the system takes its share, about 122 GiB is left for models. For comparison, an RTX 5090 has 32 GB of VRAM. On capacity alone, one Spark holds as much as four flagship gaming cards.
Its memory bandwidth, though, is only about 273 GB/s. The RTX 5090 does 1.8 TB/s, more than six times as much.
Capacity decides whether a model fits. Bandwidth decides how fast it talks.
Sorted by weight size: gpt-oss-120b takes about 61 GiB and a 4-bit GLM-4.5-Air about 68 GiB. Both fit comfortably, with more than 40 GiB left over for the context cache. Qwen3-235B only squeezes in at 3-bit, and then space gets tight. A 671B model like DeepSeek V3 needs around 380 GiB even at 4-bit, so a single Spark can't hold it.
A practical rule of thumb: keep the weights under 100 GiB. Past that line, there's no room left for a long-context cache.
Speed comes down to one formula
To generate each token, the model has to read every weight that step uses from memory. That puts a hard ceiling on single-user generation speed:
Max generation speed ≈ memory bandwidth ÷ bytes read per token
A 4-bit (Q4_K_M) Llama 70B is about 42.5 GB, and 273 ÷ 42.5 ≈ 6.4. That's its theoretical ceiling, and real-world numbers come in lower.
Ollama's official benchmarks on the DGX Spark:
At 4 tokens per second, you're watching the answer trickle out. Interactive use is painful.
The counterintuitive part: 120B runs nearly 10× faster than 70B
The most interesting row in that table is gpt-oss-120b. It has about 67% more parameters than Llama 70B, yet it generates more than 9 times faster.
The reason is architecture. Llama 70B is a dense model: every one of its 70 billion parameters takes part in every token, so all of them are read from memory each time. gpt-oss-120b is a mixture-of-experts (MoE) model: its 117 billion parameters are split into many "experts," and each token only calls on a few of them. Only about 5.1 billion parameters are active per token.
The dense 70B reads about 42 GB per token. The MoE 120B reads only about 5 GB: roughly 2 GB for the selected experts, plus about 3 GB for the attention and output layers, which are kept at higher precision. That explains the nearly tenfold speed difference on a bandwidth-limited machine.
This also sets the rule for choosing models: the Spark is a machine built for MoE. Total parameters can be large, because there's plenty of memory. Active parameters must be small, because there isn't much bandwidth.
What it can actually run
By that logic, common open-weight models fall into seven tiers:
The inference engine matters too. Forum benchmarks of gpt-oss-120b show about 46 tokens/s on llama.cpp, about 52 on SGLang, and up to 60 tokens/s on a tuned vLLM. For a single user, llama.cpp is the least hassle. For shared use with many concurrent requests, go with vLLM.
Linking two: 256 GB
The Spark has a built-in ConnectX-7 high-speed network card. Two units can be connected directly and use tensor parallelism to split one model in half, each running one half.
With 256 GB across two units, Qwen3-235B runs in full at 4-bit. In forum benchmarks, gpt-oss-120b goes from 60 tokens/s on one Spark to 72 on two, a modest gain, since it never lacked memory in the first place. DeepSeek 671B still only fits as an extreme ~2-bit quantization on two units; the quality loss is noticeable, and it's hardly worth it.
Is it worth the price?
This machine has only gotten more expensive over the past year. The NVIDIA Founders Edition with 128 GB launched at $3,999, rose to $4,699 in February 2026, and jumped to $6,950 in October, up 74% in a year. When it raised the price in February, NVIDIA cited tight memory supply; for the October increase, media coverage likewise pointed to the global memory shortage.
Now put those prices next to the bandwidth numbers we keep coming back to:
At the same 128 GB, an AMD Strix Halo mini PC has about the same bandwidth as the Spark for roughly half the price. A Mac Studio M5 Max now costs less than the Spark and has 2.2 times its bandwidth, so it generates noticeably faster. If all you want is to run big models locally, the Spark offers no clear price advantage.
What the extra money buys is NVIDIA's software stack: CUDA, vLLM, SGLang and TensorRT-LLM work out of the box, long prompts are processed quickly, fine-tuning works, and whatever you build here moves unchanged to NVIDIA servers in the data center. If you don't need those, it isn't worth the price.
Who it's for
Think of this machine as a workstation that can hold big models but doesn't talk fast:
For everyday chat and coding help, gpt-oss-120b is fast enough and good enough;For agents and batch document processing, use a smaller MoE model; in the benchmarks above, gpt-oss-20b decodes about 40% faster than gpt-oss-120b;Fine-tuning models up to 70B (QLoRA) is where 128 GB of memory really pays off;Dense 70B models only suit offline jobs where waiting doesn't matter; if you must use one, add a small draft model for speculative decoding, which speeds things up noticeably.
If what you mainly want is "one person chatting with one big model, fast," a Mac Studio or a GPU with lots of VRAM is a better fit. If you just want to hold a 100B-class model as cheaply as possible, an AMD mini PC is the better deal. The Spark is worth its price only if your work depends on CUDA, involves fine-tuning, or needs to run the same software stack as the data center.
When picking a model, ask just two questions: Do the total parameters fit in 122 GiB? Are the active parameters small enough? If both answers are yes, it's a good match for this machine.
Buy hardware, or buy a subscription?
After the hardware math, there's a more direct calculation.
My own choice is subscriptions: Claude Max 5x at $100 a month, plus ChatGPT Plus (which includes Codex) at $20 a month, for $1,440 a year. A $6,950 Spark would pay for both for nearly 5 years; for Max 5x alone, nearly 6 years.
Here's my usage on one day: 25 agent sessions, about 60% of the weekly Claude Code allowance used with 3 days to go before the reset, and a bit over half of the weekly Codex allowance. At that pace, the two subscriptions generally cover my needs, and they run frontier models. The best local option, gpt-oss-120b, still falls short of them on coding and long-running tasks.
I'm not the only one doing this math. On November 8, a Chinese-language post went viral on X: someone thought a $200-a-month AI subscription was too expensive, so they turned around and spent more than 100,000 (presumably yuan) on two Mac Studios just to run open-source models, only to find, after a lot of tinkering, that it "still isn't as good as the $200-a-month closed model." The poster's summary: "He wanted to save on subscription fees and ended up prepaying the next decade-plus of them in one go." Elon Musk quote-posted it with just three words: "Many such stories." The quote post drew nearly 5 million views.
OpenAI's Sam Altman recently explained the logic behind this on a podcast: memory in personal computers is in use only about 5% of the time (his own estimate), while the cloud, by pooling requests and batching them, can usually push utilization higher and spread the cost. He sells cloud subscriptions, so he has a stake in the argument, but the logic holds: in the cloud, one read of the weights can serve a whole batch of requests at once, spreading the cost; a personal machine serves one person at a time and sits idle much of the day. So before buying local hardware, estimate how much you'll really use it: the less you use it, the worse the deal.
Local models do offer things a subscription can't:
Privacy: client data, internal code, compliance requirements; the data can't leave your own machine;No platform restrictions: the model isn't subject to a provider's content policies, and what it answers is up to you;Offline and unlimited: it works without a connection, and batch jobs don't count against a quota;Fine-tuning: if you want to train your own model, a subscription can't do that.
For most people, a subscription is cheaper and the models are stronger. Buying a Spark only makes sense if at least one of the four needs above is non-negotiable.
#NVIDIA #LLM #artificialintelligence

Artificial Intelligence is becoming an increasingly important part of the financial industry.
And now, Binance has introduced a new development: Binance AI — a personalized AI-powered trading experience.
This is an interesting step for the crypto ecosystem.
🤖 What is Binance AI?
According to Binance's November 9 announcement, Binance AI is a new suite of features designed to provide adaptive and personalized financial experiences to users.
The idea is to bring AI-powered assistance into the Binance ecosystem.
📊 Why does this matter?
Crypto markets operate 24/7.
Traders constantly deal with:
- Market information overload
- Fast-moving price changes
- Complex trading concepts
- Research and decision-making challenges
AI tools could help users understand information more efficiently and make their research process easier.
🔍 Three potential benefits of AI in crypto
1. Better information access
AI-powered tools may help users navigate financial information and understand market developments.
2. Personalized experiences
Technology that adapts to individual needs could make financial platforms easier to use.
3. Improved learning
Beginners may benefit from AI-assisted explanations of trading terminology, market concepts and platform features.
⚠️ But remember: AI is not a magic trading machine
AI can make mistakes.
It cannot guarantee profits, predict every market movement or eliminate financial risk.
Always verify important information and make independent decisions.
💡 My perspective
I believe the combination of cryptocurrency and artificial intelligence could make financial technology more accessible.
But the real test will be how useful, transparent and reliable these tools prove to be in everyday use.
One question for the Binance community:
Would you trust an AI assistant to help you analyze the crypto market?
A) Yes, definitely.
B) Yes, but only as a research assistant.
C) No, I prefer manual analysis.
D) I want to test it first.
Let's discuss!
#BinanceAI #ArtificialIntelligence #cryptotrading #BinanceSquare $BTC $BNB $ETH

ШІ забере твою роботу в крипті: чому 90% «крипто-експертів» стануть непотрібними вже завтра? 📉🤖
Усі обговорюють графіки, відсотки ФРС та мемкоїни, але мало хто помічає справжнього чорного лебедя, який тихо підкрадається до всієї криптоіндустрії. Поки ви витрачаєте години на технічний аналіз, малювання ліній і написання постів, штучний інтелект навчається робити це в тисячу разів швидше і без емоцій.
Зніміть рожеві окуляри. Подивимося правді в очі:
1️⃣ Смерть «експертності» з Telegram. Усі ці платні канали з сигналами, аналітикою та «інсайдами» скоро підуть у минуле. ШІ аналізує блокчейн-дані в реальному часі точніше за будь-якого людину-трейдера. Залишатися посередником між ринком і новачком стає просто неможливо.
2️⃣ Алгоритми проти людських емоцій. ШІ не піддається паніці, не купує на зеленій свічці від фомо і не зливає депозит уночі від недосипу. Він холоднокровно забирає ліквідність у тих, хто покладається лише на інтуїцію.
3️⃣ Хто виживе в новій ері? Зароблятимуть не ті, хто просто тисне кнопки, а ті, хто навчився керувати технологіями та використовувати нейромережі як інструмент, а не як конкурента.
Технологічний прогрес не питатиме, чи готові ви до змін. Він просто залишить за бортом усіх, хто застряг у минулому столітті.
Питання до спільноти: Чи боїтеся ви, що штучний інтелект повністю змінить правила гри і відбере хліб у трейдерів та контент-креаторів?
🤖 — Звісно! ШІ вже дихає в спину і скоро всіх замінить.
😎 — Ніколи! Живий розум і інтуїція людини незамінні на ринку.
🤡 — Мені байдуже, я просто намагаюсь хоч щось не злити.
Пишіть свою думку в коментарях! Подивимося, скільки тут техно-скептиків. 👇💬
#artificialintelligence #TechTrends #FutureOfWork #BinanceSquare #cryptoeducation
$MSFTB
$GOOGLB
$AMDB

🤖 LA IA ESTÁ SACUDIENDO WALL STREET Y BITCOIN TAMBIÉN ESTÁ SINTIENDO SUS EFECTOS
$BTC
Muchos piensan que Bitcoin solamente depende de las noticias relacionadas con las criptomonedas.
Pero el mercado está demostrando algo diferente.
$USDT
Durante los últimos días, la preocupación por las valoraciones de las empresas tecnológicas y el gasto en infraestructura de inteligencia artificial ha contribuido a una venta generalizada de activos de riesgo.
Bitcoin también ha sentido esa presión.
¿Por qué sucede?
Los grandes inversores no siempre mantienen sus inversiones en compartimentos separados.
Cuando aumenta el temor en el mercado tecnológico, algunos reducen su exposición a diferentes activos, incluyendo criptomonedas.
Esto puede generar movimientos simultáneos en acciones tecnológicas y Bitcoin.
¿Qué podría suceder?
🟢 Escenario positivo: Las empresas tecnológicas demuestran que sus inversiones en IA están generando ingresos y recuperan la confianza del mercado.
🔴 Escenario negativo: Las expectativas de crecimiento superan los resultados reales y continúa la reducción de posiciones en activos de riesgo.
Mi análisis
La inteligencia artificial puede representar una gran oportunidad tecnológica, pero no significa que todas las empresas relacionadas con ella estén bien valoradas.
Tampoco significa que cualquier token que incluya la palabra AI sea una buena inversión.
Investigar los fundamentos es indispensable.
👇 Pregunta para la comunidad:
¿Crees que la inteligencia artificial terminará impulsando el mercado cripto o que una burbuja tecnológica podría arrastrar también a Bitcoin?
#Aİ #ArtificialIntelligence #Bitcoin #BTC #CryptoMarket #Technology