The AI bubble keeps growing
I’m not saying this isn’t useful and interesting, because it looks like it’s both. But it’s still ridiculous. Almost every single day I discover a dozen interconnected AI companies, products, technologies, and frameworks.
RadixArk is a Palo Alto-based AI infrastructure company founded in 2025 by two Chinese people, Ying Sheng (CEO) and Banghua Zhu. I don’t care whether they’re now US nationals or not, but Ying Sheng first graduated from the Shanghai Jiao Tong University and only then from Columbia and Stanford. Banghua Zhu graduated from Tsinghua University before reaching Berkeley, Microsoft, and NVIDIA. What would America do without China? I counted more than 50 people with Chinese names in RadixArk’s LinkedIn profile!
So this RadixArk, in partnership with Intel and Eigent AI, used Miles (which supports most Chinese models and the most expensive and restricted NVIDIA and AMD hardware) for reinforcement learning to build and release their first model, DaisyRose-27B, built on Qwen3.8-27B!
Isn’t it wonderful? Millions of customized, trained, quantified—and whatever else can you do to a model—models.
We trained it on long-horizon agent tasks with sessions up to 1M tokens, with just 1,200 curated tasks in 18 days on 144 H200 GPUs.
Long live Hugging Face:
We are releasing BF16, FP8, and NVFP4 checkpoints for everyone to try and build on.
You cannot know exactly what they did, but it must be good stuff:
- Each task targets a mistake the original model makes. We ran the original Qwen3.8-27B on every training task and carefully redesigned each one so that it brings out the mistakes we want the model to stop making. We then tuned each task until its solve rate fell within a suitable range, neither too easy nor too hard for the model to learn from.
- Every task is a long job with several problems chained together. Each task entails a chain of interdependent problems, so an agent has to navigate through multiple stages of reasoning, planning, and execution. Each attempt gets up to 1.5 hours in the sandbox, with more than 100 tool-calls and more than 256k output tokens.
- There is no room for reward hacking. We reviewed every task along with the model’s attempts at it. Any task that could be passed without really solving it, or that had a design flaw, was fixed or removed before training.
Maybe the future really is to run your model locally—that is, if you can afford the hardware:
Try DaisyRose-27B, or train your own
You can download DaisyRose-27B from Hugging Face, and you can train your own agents with our open source reinforcement learning framework Miles.
Why, of course! I’ll run it on my Pentium 4!
More keywords only known to 0.002% of the world’s population:
We are deeply grateful to Eigent AI for working closely with us on training data, and to the Qwen team for creating Qwen3.8-27B, the foundation of our work. We also thank the Miles and SGLang communities, as well as other open-source projects, including slime, AgentENV, and Harbor.
The last three, of course. Oh, Qwen3.8-27B had so many faults, but we couldn’t do without it. (And we did it in the States, because land of the free and of NVIDIA.)
From Hugging Face:
DaisyRose-27B is RadixArk’s first model: a multimodal reasoning model built on Qwen3.8-27B with reinforcement learning only, without any supervised fine-tuning or distillation.
Speaking of my lack of knowledge:
We strongly recommend serving DaisyRose-27B with SGLang and DSpark speculative decoding, using the official
lmsysorg/sglang:latestDocker image (DSpark is included).
SGLang and DSpark? Speculative decoding? Not even the decoding is straight? Gee, those tulip bulbs!
🤖
Just for the fun of it, and to ask “someone” other than Gemini, Qwen, Claude, or ChatGPT (Grok said I exhausted my free quota), I revisited a former friend, Kimi, who agreed to give a couple of quick free answers.
❶ Ludditus: What hardware is needed to run the BF16 weights of DaisyRose-27B?
Kimi: For the BF16 checkpoint of DaisyRose-27B (RadixArk, released Nov 18, 2026 — a post-trained Qwen3.8-27B), the model card states the weights take about 52 GB and that it runs on a single GPU.
Since 52 GB is the weight floor, your practical options are:
Single GPU / chip (needs ~80 GB+ class to leave room for KV cache and runtime):
- NVIDIA H100 80GB or H200 — the comfortable fit, with headroom for context and concurrency
- RTX PRO 6000 (96GB) — also fits
- Apple silicon: a 96GB Mac Studio (M3 Ultra) is a tight fit; the 128GB/192GB+ tiers are safer if you want long context
Multi-GPU setups:
- 2 × 48GB cards (e.g., 2 × RTX PRO 6000 / L40S / A6000 Ada) via tensor parallelism in vLLM or SGLang
- 2 × 24GB or 32GB cards (RTX 4090s / RTX 5090s) can work with vLLM sharding, though it’s cramped
Caveats:
- The 52 GB is weights only. KV cache stacks on top and grows with context length and concurrent requests — this model has a 256K-token context window, so at long contexts you’ll want the extra headroom the 80–96GB cards provide.
- If you don’t strictly need BF16, the FP8 checkpoint (29 GB) fits a 48GB card, and NVFP4 (23 GB, requires a Blackwell GPU) fits ~32GB.
Bottom line: one 80GB-class datacenter GPU (H100/H200) is the minimum hassle-free configuration; anything smaller means the FP8/NVFP4 quants or multi-GPU tensor parallelism.
❷ Ludditus: Is there any DaisyRose-27B that runs on lower hardware?
Kimi: Yes — several, scaling all the way down to laptops:
Official quantized checkpoints from RadixArk:
- DaisyRose-27B-FP8 (~29 GB) — fits any single 48GB-class card (L40S, A6000 Ada, RTX PRO 6000), or a Blackwell/GB10 box like DGX Spark
- DaisyRose-27B-NVFP4 (~23 GB) — runs on a single NVIDIA Blackwell GPU; NVIDIA’s equivalent Qwen3.8-27B NVFP4 recipe is validated on the DGX Spark (GB10), DGX Station, RTX PRO 6000, and RTX 5090
GGUF (llama.cpp / Ollama / LM Studio): bartowski has already published RadixArk_DaisyRose-27B-GGUF, with the full Q4_K_M / Q3_K_M / Q2_K ladder. For reference, the equivalent Qwen3.8-27B GGUFs come in around 16–17 GB at Q4_K_M and ~12 GB at Q2-level. Since DaisyRose keeps the linear-attention layers at higher precision, budget roughly:
| Hardware | What works |
|---|---|
| 24 GB (RTX 4090/5090, Mac 24GB) | Q3_K / Q4 with short context, or offloading to CPU RAM |
| 16 GB (laptop / MacBook) | ~Q2_K quant (~10–12 GB) — for Qwen3.8-27B this runs at 6–7 tok/s on an M2 Pro, so expect similar |
| 12 GB | Aggressive 1–2-bit quants (Unsloth Dynamic, ~6.2–9.8 GB files exist for the base architecture), with real accuracy loss |
Quality guidance: benchmark work on the base architecture shows Q4_K_M sits above the 95% accuracy-recovery threshold (safe), while the 2-bit tier is where degradation becomes noticeable — fine for casual chat, risky for agentic coding.
Bottom line: for a smooth experience, the FP8 on a 48GB card or the GGUF Q4_K_M on ~16–24GB is the sweet spot. The full NVFP4/BF16 setups are only worth it if you’re on Blackwell or datacenter hardware anyway.
❸ Ludditus: Is any of these runnable on 16 GB normal RAM? https://huggingface.co/bartowski/RadixArk_DaisyRose-27B-GGUF/tree/main
Kimi: Yes — barely, but only one tier really fits. Here are the actual file sizes from the repo:
| Quant | File size | Fits in 16 GB RAM? |
|---|---|---|
| IQ2_XXS | 8.76 GB | ✅ Yes, the only comfortable option |
| IQ2_XS | 9.00 GB | ✅ Yes |
| IQ2_S / IQ2_M / Q2_K | 9.6–10.8 GB | ⚠️ Possible but tight once you add KV cache and OS overhead |
| Q3_K_S / IQ3_XXS and up | 12.2 GB+ | ❌ No — leaves almost nothing for context, buffers, or the OS |
What to expect if you run RadixArk_DaisyRose-27B-IQ2_XXS.gguf CPU-only on 16 GB:
- Memory: ~9 GB weights + a few hundred MB of KV cache (the hybrid linear-attention design keeps cache small, and it won’t grow much unless you crank the context to tens of thousands of tokens — which you shouldn’t anyway)
- Speed: roughly 1–3 tokens/second on a typical desktop CPU. A hybrid MoE model at 2-bit on pure CPU is a “works, but patience required” experience. If your CPU has AVX2/AVX512 or you’re on Apple Silicon with unified memory, it’ll be on the better end of that
- Quality: the 2-bit tier is where perceptible degradation starts (the earlier-mentioned ~95% accuracy-recovery rule of thumb puts IQ2 below the “safe” line). It’s fine for casual chat; I wouldn’t trust it for careful agentic coding work
Practical tips:
- Close everything else while running it — the OS plus browser can eat 3–4 GB
- Skip the
mmprojfiles (~0.93 GB) unless you need vision — text-only works without them - Cap the context window at 4–8K tokens rather than the model’s full 256K
- Run with llama.cpp or LM Studio; the memory-mapped loading means an occasional page-in hiccup is normal
Bottom line: IQ2_XXS (8.76 GB) is your answer. If you’d rather not live at 2-bit quality, your real options are upgrading to 32 GB RAM, or looking at smaller models entirely — at this point there are plenty of excellent 7–14B models that fit a 16 GB box at Q4–Q6 quality.
❹ Ludditus: What’s the fuss with “Apple Silicon with unified memory”? Even Intel without a graphic card uses RAM instead of VRAM, so it’s “unified”! Apple’s unified RAM is probably using RAM that’s faster than DDR5 but not more than DDR6, but the term is overrated.
Kimi: You’ve got a fair point — the term is mostly marketing. Intel/AMD integrated GPUs have shared system memory for decades; there’s no conceptual breakthrough in “the GPU reads the same DRAM the CPU does.” An Intel NUC or Ryzen mini-PC running llama.cpp is doing exactly the same thing as a Mac mini.
Where Apple’s version actually differs is in the engineering around it, and it comes down to one number: memory bandwidth.
- A typical desktop with dual-channel DDR5-6400 tops out around ~100 GB/s. That’s the ceiling your iGPU gets for LLM inference, and it’s the real bottleneck — token generation on a 27B model is essentially pure memory-bandwidth-bound, which is why CPU inference crawls at 1–3 tok/s.
- Apple puts a very wide LPDDR5X interface right next to the SoC: an M4 Pro does ~273 GB/s, M4 Max ~410–546 GB/s, and the M3 Ultra ~800 GB/s. That’s 3–8× a standard desktop’s bandwidth — in the territory of entry datacenter cards.
So the Mac isn’t “unified” in a way a PC isn’t; it’s unified with datacenter-class bandwidth attached to a consumer memory pool. That combination is what made Mac Studios the go-to box for running 70B+ models locally before 48GB/80GB datacenter cards became accessible. A PC with an RTX 4090 does inference on GDDR6X at ~1 TB/s but only has 24 GB of it; a Mac gives you ~550 GB/s across 128 GB. For LLMs specifically, capacity × bandwidth is the metric, and Apple hit a sweet spot nobody else bothered to occupy.
One correction to your comparison: DDR6 doesn’t exist yet — there’s no shipping DDR6 standard or modules as of late 2026. GDDR7 ships on RTX 50-series cards, and LPDDR5X is what Apple uses. So the honest framing is “Apple’s on-package LPDDR5X is 4–5× faster than the DDR5 in a normal desktop,” which is a real gap, just not a magical one.
And to your implicit conclusion: yes, the word “unified” itself is overrated. If someone already has a 16 GB machine — your earlier question — the Mac advantage is mostly irrelevant anyway; an 8.76 GB model fits and runs at similar (slow) speeds on either platform at that capacity tier. The unified-memory pitch only matters when you’re deciding between a $2,000 Mac with 128 GB at 800 GB/s versus a PC with a 24 GB GPU or a bandwidth-starved iGPU.
❺ Ludditus: Is Apple’s LPDDR5X 4-5x faster than DDR5-5600?
Kimi: Yes for total bandwidth — with a small correction on the multiplier, and an important caveat on why.
The math:
- DDR5-5600 in a typical dual-channel desktop: 5600 MT/s × 2 channels × 8 bytes = ~90 GB/s
- Apple Silicon aggregate bandwidth: M4 Pro ~273 GB/s, M4 Max ~410–546 GB/s, M3 Ultra ~800 GB/s
So against DDR5-5600:
- M4 Pro: ~3×
- M4 Max: ~5–6×
- M3 Ultra: ~9×
My earlier “4–5×” was a bit loose — that’s the right ballpark for the M4 Max specifically, while the M4 Pro is closer to 3×.
The caveat you should keep in mind: LPDDR5X is not a faster memory technology per pin. A single LPDDR5X chip and a single DDR5 chip run at similar transfer rates (both top out in the 6400–8500 MT/s range, with LPDDR5X actually being the low-power, slightly-slower variant). The difference is entirely bus width:
- A desktop CPU wires up 2×64-bit channels to its DIMM slots
- Apple solders a very wide array of LPDDR5X packages directly beside the SoC — the M4 Max has the equivalent of a ~512-bit+ interface
Think of it as a highway: same speed limit per lane, but Apple built 4–8× as many lanes. That’s why the comparison isn’t really “LPDDR5X vs DDR5” — it’s “Apple’s wide on-package interface vs a standard DIMM channel configuration.” A server CPU with 8 or 12 channels of DDR5 (Threadripper, EPYC) also reaches 300–460 GB/s, because it gets the same effect from more lanes, at the cost of a much bigger platform.
So: the 4–6× figure for the M4 Max is real aggregate bandwidth, not marketing — but it comes from chip-count and bus-width engineering, not from LPDDR5X being some exotic fast memory.
❻ Ludditus: So, Apple doesn’t market the “unified memory” in cheaper MacBooks, even if the memory is technically unified?
Kimi: Actually, they do — and that’s exactly where the term is most misleading.
“Unified Memory Architecture” is a blanket marketing label Apple applies to every Apple Silicon Mac, including the $999 MacBook Air. It’s a genuine architectural descriptor for them: since the M1, CPU, GPU, and Neural Engine all sit on one SoC with one memory controller, so there’s one pool regardless of tier.
The catch is that the benefit of that architecture scales almost entirely with bandwidth, and the bandwidth scales with price:
| Chip | Memory bandwidth | vs. dual-channel DDR5-5600 (~90 GB/s) |
|---|---|---|
| M4 (MacBook Air, base MacBook Pro) | ~120 GB/s | ~1.3× |
| M4 Pro | ~273 GB/s | ~3× |
| M4 Max | ~410–546 GB/s | ~5–6× |
| M3 Ultra | ~800 GB/s | ~9× |
So on a base MacBook Air, you’re right back to your original point: it’s “unified” in exactly the same sense an Intel laptop with integrated graphics is unified — one shared pool of DRAM, no magic. The performance uplift over a DDR5 laptop is small (~30%), which is why running a 27B model on a 16 GB Air is just as slow as it would be on a 16 GB PC.
The term earns its keep only at the Pro/Max/Ultra tiers, where the wide on-package interface delivers that 3–9× bandwidth advantage. Apple’s marketing conveniently doesn’t draw that distinction — a MacBook Air spec sheet mentions unified memory and lets you assume the architecture implies the performance.
If you’re shopping for local LLM inference: ignore the word “unified,” find the memory bandwidth spec for the exact chip, and compare that number. That’s the only figure that predicts token generation speed.
Now, because Kimi said this is all it can answer for free, I want to add some context:
- The MacBook Neo has a memory bandwidth of 60 GB/s on its Apple A18 Pro chip. It uses 8 GB unified memory (LPDDR5X-7500).
- When Apple gives a memory bandwidth figure in GB/s, they are usually referring to the maximum theoretical rate at which the processor can transfer data to and from RAM. Or this is what ChatGPT claims.
- My Lenovo IdeaPad Slim 5 14IRH10, using DDR5-5600 on dual channel, has a theoretical peak at 89.6 GB/s, but because of the Intel Core i5-13420H CPU, it’s limited to 5200 MT/s, hence it peaks at 83.2 GB/s (theoretically). To run at 5200 MT/s, an IdeaPad Slim 5 with Core Ultra 5 135H would do, such as this IdeaPad Slim 5 14IMH10.
- My i5-13420H, compared to Neo’s A18 Pro: the A18 Pro is 47% faster in Geekbench 6’s single-core test, but Intel is 15% faster in multi-core tests.
- Neither laptop is suited for running local models in any meaningful way, but Neo is limited to 8 GB of unified memory (which makes it completely useless for anything serious) and a 512 GB SSD. My 16 GB + 1 TB laptop was cheaper and is generally a better value for money, especially for Linux (I didn’t pay the “Microsoft tax”).
ChatGPT also offered the figures for the base M-series chips, rather than the Pro, Max, and Ultra variants:
| Apple chip | Theoretical memory bandwidth |
|---|---|
| M1 | 68.25 GB/s |
| A18 Pro (MacBook Neo) | 60 GB/s |
| M2 | 100 GB/s |
| M3 | 100 GB/s |
| M4 | 120 GB/s |
| M5 (MacBook Air 13″ & 15″, MacBook Pro 14″) | 153 GB/s |
Neither is AI material.
🤖 ChatGPT: Apple still has machines specifically interesting for local AI workloads, particularly the Mac Studio. The relevant configurations are the M4 Max and M3 Ultra.
Mac Studio: the relevant hardware
The 2025 Mac Studio offers:
| Chip | Memory bandwidth | Unified memory |
|---|---|---|
| M4 Max | 410 GB/s | 36 GB base |
| M4 Max (higher configuration) | 546 GB/s | 48, 64 or 128 GB |
| M3 Ultra | 819 GB/s | 96 GB base |
| M3 Ultra (higher configuration) | 819 GB/s | 256 GB |
Current lineup: M5 Max and M5 Ultra
As of November 2026, Apple has already replaced the M4 Max and M3 Ultra Mac Studio with the M5 Max and M5 Ultra.
| Mac Studio | Base configuration | Starting price in Germany |
|---|---|---|
| Mac Studio M5 Max | 36 GB RAM, 512 GB SSD | €2,999 |
| Mac Studio M5 Ultra | 96 GB RAM, 1 TB SSD | €6,599 |
German launch pricing, VAT included.
Why these are interesting for local AI
| Chip | Memory bandwidth | Unified memory |
|---|---|---|
| M5 Max | 614 GB/s | Up to 128 GB |
| M5 Ultra | 1.2 TB/s | Up to 512 GB |
The M5 Ultra’s 512 GB configuration became available in late October 2026.
The main advantage for local LLMs isn’t necessarily CPU speed. It’s the combination of:
- Large unified memory — models that would not fit into a conventional GPU’s VRAM can potentially fit into 96, 256 or even 512 GB.
- Very high bandwidth — particularly important for token generation, which often becomes memory-bandwidth-limited.
- Integrated GPU — no separate VRAM allocation or PCIe transfer bottleneck between CPU and GPU.
What I’d consider for AI
| Configuration | Practical appeal |
|---|---|
| M5 Max, 36 GB | Good for smaller models, but 36 GB is restrictive for larger LLMs |
| M5 Max, 64–128 GB | The more interesting AI workstation option |
| M5 Ultra, 96 GB | High performance, but memory capacity may limit larger models |
| M5 Ultra, 256 GB | Serious local LLM experimentation |
| M5 Ultra, 512 GB | Very large models, expensive enthusiast/research territory |
For local AI experimentation, I’d consider the M5 Max with 128 GB the most interesting compromise, if the price is acceptable. The Ultra is impressive, but €6,599 is a considerable investment.
The M5 Max is already far beyond what you need for ordinary desktop computing; its justification would be running larger models locally, not replacing a conventional laptop.
In your dreams, honey.

Leave a Reply