the short version
- The largest measured change is prefill, not decode: prompt processing is up 150% on average over the M3 Ultra, a ~2.5x improvement, which is what makes agent turns with large system prompts usable.
- Memory bandwidth went from 819 GB/s to 1.2 TB/s, a 50% increase. The unified memory ceiling did not move: still 512 GB, with that configuration shipping in late October.
- On a dense Qwen3.8 27B, an RTX 5090 generates faster than the M5 Ultra at every context length both machines ran. The Mac's argument is models that do not fit in 32 GB of VRAM at all.
- The review reports no time-to-first-token figures in milliseconds, no batched-serving numbers, and no dense-model prompt processing figure for the 5090.
MacStories published a four-day hands-on of the M5 Ultra Mac Studio on 21 September 2026, and the number that matters for local serving is prefill: prompt processing is up 150% on average over an M3 Ultra, a ~2.5x improvement, while token generation is about 70% faster on average in side-by-side tests through Open Minis for iOS. Memory bandwidth rose from 819 GB/s to 1.2 TB/s, 50% higher, and the unified memory ceiling did not move - it is still 512 GB, with that configuration shipping in late October. The unit tested had 256 GB. For anyone serving models locally, the change is in which workloads finish fast enough to sit inside an agent loop, not in peak decode throughput.
The comparison covers three machines: an M5 Ultra Mac Studio with 256 GB of RAM, an M3 Ultra with 512 GB, and a desktop PC with an RTX 5090. The reviewer states he is not an AI developer by trade and does not train or fine-tune models, so the figures come from inference under working agent harnesses rather than synthetic kernel benchmarks. The models are MLX models served through oMLX, driving Open Minis for iOS, Hermes Agent and the Codex app.
Bandwidth up 50%, the 512 GB ceiling unchanged
The M5 Ultra uses UltraFusion to connect two dual-die M5 Max chips into a quad-die architecture, which the review calls a first for the Apple ecosystem. The GPU has 80 cores, each with a Neural Accelerator, and the review reports up to 4.5x the peak GPU compute for AI compared to the M3 Ultra. The measured gains do not track either axis cleanly: prompt processing moved 150%, well past the 50% bandwidth increase and well short of the 4.5x compute figure, while generation moved about 70%.
The capacity story is the one that did not change. Apple's unified memory architecture still tops out at 512 GB, as it did on the M3 Ultra. The tested 256 GB machine ran a 5-bit quant of Qwen3.8-Flash-Next entirely in RAM; 6-bit and 8-bit quants ran only by offloading their n-gram tables to SSD. The reviewer settles on 5-bit as the balance of intelligence, performance and memory consumption, and says he would want to measure the 8-bit quant without SSD offloading on a 512 GB machine - a measurement that does not exist yet.
Prefill gains matter more than decode for agents
In the review's account, the blocker for local models in agent harnesses was not steady-state tokens per second. A modern agent turn opens with a system prompt, user personalization and session memories, skill and MCP descriptions and more, so the context window is never empty; the reviewer describes staring at a loading indicator before any token appeared, with every turn of the loop getting worse as session context grew. The 2.5x cut in prompt processing is what changed that, and Qwen3.8-Flash-Next is now the default model in both Open Minis for iOS and Hermes Agent. On a longer project, the review reports the model processing multi-turn loops, dispatching and coordinating subagents, and doing it at 60 to 85 tokens per second as the thread grows.
with the M5 Ultra, you spend considerably less time waiting for a model to read your prompt and begin generating a response
The RTX 5090 still wins on dense 27B
A commenter on Hacker News transcribed the dense-model chart from the review, which compares generation speed on Qwen3.8 27B across prompt sizes. At 8K the RTX 5090 does 59 tokens/sec, the M5 Ultra 48, the M3 Ultra 31. At 64K: 51, 39, 23.5. At 128K: 44, 32, 20. At 256K the 5090 has no figure, while the M5 Ultra does 24 and the M3 Ultra 15. The review concedes the point directly: a 5090 still has an edge over the M5 Ultra thanks to its higher memory bandwidth.
On a dense model that fits in 32 GB of VRAM, the Nvidia box wins at every context length both machines ran. The M3 Ultra to M5 Ultra gain on those same dense numbers is 55% at 8K and between 60% and 66% at 64K through 256K, closer to the 50% bandwidth increase than to the 4.5x compute claim. The review publishes no dense-model prompt processing number for the 5090, so the 150% prefill gain is an M5-versus-M3 figure only and the cross-vendor prefill comparison is not settled by this data.
Why MoE, not dense 27B, suits the Mac
A dense 27B doesn't really make sense for the Mac. A MoE makes way more sense when you have modest bandwidth but lots of memory.
The review's Flash-Next numbers are higher than its dense 27B numbers: Qwen3.8-Flash-Next clears 100 tokens/second on short prompts and still writes at 60 to 85 tokens/second with 64K to 256K of context behind it. The review does not state Flash-Next's architecture or parameter count, so the MoE framing belongs to the commenters, not the reviewer. Another commenter makes the corollary point: run Qwen3.8-Flash-Next on a 5090 and throughput is near zero, because the model does not fit in the card's 32 GB of VRAM. The same thread argues Nvidia segments on VRAM deliberately - an RTX 6000 Pro with 96 GB has the same memory bandwidth as a 5090 and roughly 24k versus 21k CUDA units, and has soared beyond $15,000 from $10,000-11,000 three to six months ago.
Capacity also buys concurrency. With 256 GB, the review reports stacking up to three concurrent Flash-Next sessions with subagents in oMLX, which is what makes a main GPT-6 Astra thread orchestrating local subagents in Codex plausible rather than theoretical. The serving path is unremarkable in a good way: oMLX exposes an OpenAI-compatible API locally, fronted by a small local server and reached from an iPhone over Tailscale.
Image generation lands in the same memory-first bucket. The wallpaper in the review was generated by Qwen-Image-2.1 locally on the M5 Ultra in 180 seconds, with peak RAM usage of 78 GB - more than double the 32 GB of VRAM on the 5090 that commenters cite. The review reports no comparison figure for the same generation on the RTX 5090.
Practitioners push back on missing speculative decoding
The loudest objection in the Hacker News thread is that the published numbers are baseline decoding with no speculative techniques. One commenter reports a 27B doing 75 tok/s at 256k context and starting at 80 tok/s at 8k on an M5 Max with the z-lab dflash-2 collection applied, against roughly 30 tok/s at iq4 without it, and argues it is not fair to judge hardware without MTP or dflash enabled. Another posts llama.cpp figures on a 500W RTX PRO 6000: pp512 3920 tok/s and tg128 81 tok/s on Qwen 3.8 27B UD-Q4_K_M, rising to 140 tok/s with MTP3 on mtp-bench; on vLLM with NVIDIA/Qwen3.8-27B-NVFP4, prefill is about 13,000 tok/s and decode goes from a 72 tok/s baseline to 157 at mtp7 and 215 with dflash7. A third reports 2x AMD R9700, at 644 GB/s each, running Qwen3.8-27B-MXFP4 (W4A8) at 260+ tok/s falling to ~110 tok/s over 128k context at batch size 1, using speculative decoding, compressed inter-GPU all-reduce and custom GEMM kernels.
Three other threads of pushback are worth carrying into a purchase decision. One commenter wants Apple to target time-to-first-token specifically, saying that coming from an M3 Pro it still looks laggard compared to Nvidia's TensorCores in the 5090. Another assumes the Mac numbers are non-batched and estimates 4x to 8x headroom if queries can be batched, depending on model quant; the review does not test batching either way. On power, one commenter notes the 5090 draws 600W or more while the Mac is probably half that at most, where the review makes only a qualitative argument about size, heat and noise. One commenter also pulls ~30 tokens/second on GLM 5.3-flash out of the same charts.
Cost is where the thread is most skeptical. Commenters put the 512 GB upgrade at an additional $4,000-6,000 and an Ultra with 2 TB of storage north of $15,000, which one frames as about 12 years of OpenAI Pro; another works it as five years of $200/month tokens against an RTX 5090 PC at roughly half the price. One argues plainly that anyone buying local hardware to save money on models that are remotely capable is likely to be disappointed. The review's counter is narrow and concrete: agents based on DeepSeek V4 Flash, plus olmOCR for PDFs, ran 24/7 for 99 days across 310 documents for a total cost of $0, a persistent always-on background workload the reviewer judged cost-prohibitive on the OpenAI or Anthropic APIs. For interactive work the review prices the cloud alternatives too - Inco serving Kimi K3 at over 300 TPS, where $20 of credits went in 10 minutes, Cerebras running Qwen3.8-27B at 1,800 TPS, and Fireworks and Baseten each above 150 TPS.
What these benchmarks do not cover
The 512 GB M5 Ultra has not been benchmarked by anyone in this material; the review's own open question is 8-bit Flash-Next without SSD offloading, and that configuration ships in late October. There are no batched-serving numbers, no time-to-first-token figures in milliseconds for any of the three machines, and no measurement of how the n-gram table SSD offload path behaves under sustained load. The comparison that would settle the argument - Flash-Next on the M5 Ultra against a comparable model on an RTX 6000 Pro with speculative decoding enabled on both sides - has not been run. Further out, commenters point to a Bloomberg report by Mark Gurman, via Tom's Hardware, that an M7 Ultra planned for 2028 would support up to 1.5 TB of unified memory, with a base M6 this fall, a base M7 in the first half of 2027 and M7 Pro and Max at the end of 2027; that is reporting on a roadmap, not a specification.
Questions this raises
how much faster is the M5 Ultra than the M3 Ultra for local LLMs
Prompt processing is up 150% on average, roughly a 2.5x improvement, and token generation is about 70% faster in side-by-side tests. The gains sit between the 50% memory bandwidth increase and the claimed 4.5x peak GPU compute for AI.
does the M5 Ultra Mac Studio support more than 512 GB of unified memory
No. The unified memory ceiling is unchanged at 512 GB, the same as the M3 Ultra, with that configuration shipping in late October. The tested machine had 256 GB, which ran a 5-bit quant of Qwen3.8-Flash-Next entirely in RAM while 6-bit and 8-bit quants needed SSD offloading for their n-gram tables.
is an RTX 5090 still faster than an M5 Ultra for local models
On a dense Qwen3.8 27B that fits in 32 GB of VRAM, the 5090 wins at every context length both machines ran, and the review credits its higher memory bandwidth. But commenters note that running Qwen3.8-Flash-Next on a 5090 gives near zero throughput because the model does not fit in 32 GB.
These daily notes are drafted by a model I run and operate myself - the same kind of pipeline this site is about - from sources published in the previous 24 hours, and every one lists what it read. The longer essays, the talks and the preprint are mine, written by hand.