Apple and Meta have enabled Meta's 816-billion-parameter Muse Spark Open model to run locally on clustered Mac Studios. CNBC reported that each 256GB M5 Ultra machine costs at least $10,799, so a two-machine cluster comes to $21,598 or more.
Perspective Coverage
31 publishers
- Builder
- Builder 41%
- Operator
- Operator 28%
- Investor
- Investor 31%
Reality
- Evidence55
- Adoption30
- Hype gap+30
- Incentives65
- Confidence68
Apple and Meta get Muse Spark Open running on two to four Mac Studios
Apple says four Mac Studios run Meta's Muse Spark Open at over 36 tokens per second, Computerworld reports. The speed is Apple's own measurement, and the cost case is the columnist's inference from the price of four machines.
Reality
- Evidence35
- Adoption
- Not scored
- Hype gap+40
- Incentives70
- Confidence55
Four Mac Studios linked over Thunderbolt run an 816B-parameter model at over 36 tokens a second, Apple says
Apple says a cluster of four Mac Studios linked with RDMA over Thunderbolt generates over 36 tokens per second from an 816-billion-parameter model. The figure is Apple's own, from its MLX LM benchmark tool, and it covers generation speed only.
Publishers:machinelearning.apple.com
Reality
- Evidence45
- Adoption
- Not scored
- Hype gap+25
- Incentives75
- Confidence55
buildOne report1 publisher Ollama's timings block rates a one-token prefill at 3,875 tokens per second
Ollama's OpenAI-compatible endpoint reported 3,875 prompt tokens per second where the real rate was 125, a dev.to author found. Any benchmark that repeats a prompt and reads that field overstates prefill speed by the share served from cache.
Reality
- Evidence55
- Adoption
- Not scored
- Hype gap+5
- Incentives35
- Confidence55
Apple restored the 512GB configuration it spent last year deleting, and put the RAM shortage on the Mac mini's sticker instead. The ceiling is the product. The ceiling is also what disappears.
Perspective Coverage
28 publishers
- Builder
- Builder 30%
- Operator
- Operator 48%
- Investor
- Investor 22%
Reality
- Evidence62
- Adoption30
- Hype gap+28
- Incentives58
- Confidence64
buildConfirmed6 publishers Alibaba's new 27B open-weight model defaults to its deepest reasoning setting. One local run burned 22,276 reasoning tokens to produce 3,223 tokens of output, and took 21 minutes.
Perspective Coverage
6 publishers
- Builder
- Builder 47%
- Operator
- Operator 39%
- Investor
- Investor 14%
Reality
- Evidence55
- Adoption25
- Hype gap+30
- Incentives55
- Confidence60
buildOne report1 publisher Conway plans to fund its on-device Underdog assistant with a cut of the purchases it handles
Conway Research's Underdog runs a 27-billion-parameter assistant model on users' own computers and plans to fund itself with a cut of purchases. Its revenue depends on how often users trust it to complete a checkout.
Reality
- Evidence40
- Adoption
- Not scored
- Hype gap+30
- Incentives60
- Confidence45
buildOne report1 publisher Open 0.8B model Jeff answers agents' multiple-choice calls and escalates the unsure ones to a 27B
Jeff, an open 0.8B model, returns a probability for each option in one forward pass and sends low-confidence agent decisions to Qwen3.8-27B. The confidence score attached to each answer tells an agent when a small decision is worth the larger model's time.
Reality
- Evidence38
- Adoption15
- Hype gap+15
- Incentives
- Not scored
- Confidence35
buildOne report1 publisher Community MLX port makes Qwen3-TTS the fastest offline voice cloner on an M1 Pro
Qwen3-TTS went from RTF 9 under PyTorch to faster than real time through a community MLX port on one developer's M1 Pro. The port took it from written off to first place among the commercially usable cloners tested for an offline Mac app.
Reality
- Evidence40
- Adoption
- Not scored
- Hype gap+25
- Incentives20
- Confidence40
buildConfirmed7 publishers The M5 Ultra is pitched as an alternative to CUDA workstations, but the 512GB configuration that carries the argument ships in late October with no announced price.
Perspective Coverage
7 publishers
- Builder
- Builder 45%
- Operator
- Operator 31%
- Investor
- Investor 24%
Reality
- Evidence60
- Adoption20
- Hype gap+35
- Incentives70
- Confidence65
Apple's M5 Ultra Mac Studio starts at $5,499 and reached $12,299 as tested, and the reviewer running local agents on it every day says the most expensive Anthropic subscription would still have cost less for years.
Perspective Coverage
10 publishers
- Builder
- Builder 36%
- Operator
- Operator 38%
- Investor
- Investor 26%
Reality
- Evidence82
- Adoption41
- Hype gap+18
- Incentives58
- Confidence76
Bonsai 2 27B keeps 98 percent of Qwen3.8 27B's aggregate benchmark score, up from 95 percent for PrismML's first release, and TechCrunch reports the 5.9GB file is small enough for a PC and possibly a high-end phone.
Reality
- Evidence44
- Adoption38
- Hype gap+33
- Incentives71
- Confidence55
buildConfirmed2 publishers PrismML's ternary build of Qwen3.8 27B keeps 98.2 percent of the full-precision benchmark average on both of the company's inconsistent scorecards, and the loss it does take is concentrated in knowledge and reasoning.
Reality
- Evidence38
- Adoption20
- Hype gap+32
- Incentives72
- Confidence55
buildOne report1 publisher Identical 0.95 confidence on right and wrong answers killed Scribe's threshold gate
Nine local VLM configurations read the same 23 handwritten clinical pages against 221 hand-curated gold values, and accuracy on the fields accepted without human review rose from 75% to 96% once a second model started checking the first.
Reality
- Evidence58
- Adoption8
- Hype gap−10
- Incentives45
- Confidence58
buildOne report1 publisher Running llama-server puts context, KV cache and GPU placement in your command line
A dev.to guide puts the llama.cpp and Ollama decision on ownership. Under Ollama the named model is the unit you operate; under llama.cpp it is the llama-server process and the seven flags in its command line.
Reality
- Evidence52
- Adoption
- Not scored
- Hype gap0
- Incentives30
- Confidence58
buildConfirmed2 publishers Edge0's Apache-2.0 engine reports a 35B model peaking at 2.9 GB of RAM on a 24 GB Mac mini. Work the decode rate against the active set and most of those expert reads have to be coming from the operating system's file cache.
Reality
- Evidence38
- Adoption12
- Hype gap+48
- Incentives55
- Confidence52
Apple's quad-die M5 Ultra puts 512GB of inference memory under the desk
Apple says the M5 Ultra reaches 512GB of unified memory at 1.2TB/s, which settles how big a local agent can get, though the September launch arrives without prices and every AI speed figure is measured against an M1.
Reality
- Evidence32
- Adoption20
- Hype gap+45
- Incentives60
- Confidence45
buildOne report1 publisher RamaLama ships models as OCI images you can inspect and sign
Treating the runtime and the weights as a container image buys you a hardened default docker run and a signable artifact. On Apple Silicon, though, that same boundary costs you the GPU. One hands-on run puts a number on that cost.
Reality
- Evidence52
- Adoption20
- Hype gap−5
- Incentives30
- Confidence48
buildOne report1 publisher Counting scales per group of 64 prices a 276B model on a 24GB Mac mini
A dependency-free script turns config.json into byte counts, checks itself against two artifacts its author did not build, and refuses to estimate anything when a check fails. The second check stopped passing in August, and the reason is worth the read.
Reality
- Evidence58
- Adoption12
- Hype gap−14
- Incentives22
- Confidence52
buildOne report1 publisher Mac Studio M5 Ultra vs DGX Spark: capacity says what fits, bandwidth says what you wait for
The bandwidth gap is 4.4x and the capacity gap is 4x, which is why these two boxes are not really competing. One decides whether a model fits; the other decides whether it is usable.
Reality
- Evidence38
- Adoption32
- Hype gap+25
- Incentives42
- Confidence35