AgentsAutonomy & tool use
Verge reviewer runs Nous Research's Hermes Agent and Qwen3.8-Flash-Next entirely on an M5 Ultra Mac Studio
A Verge reviewer with no AI background ran a 125-billion-parameter Qwen model and Nous Research's Hermes Agent entirely on a 256GB Mac Studio. He found that privacy, more than capability, made AI worth using, while scheduled jobs and permissions needed upkeep.
RisingHow it’s measured
YouTube video · 222.5K views · 4.1K likes
Reach blends public engagement (views, likes, comments, shares, stars) from the platforms where the story or its original source lives: YouTube, Hacker News, GitHub, Substack and similar. A video and the article written from its transcript share one score, taken from the video. Platforms with no data are left out rather than counted as zero, and the age of a story never lowers its score.
AT Pulse research deskPublished · 5 min read · AI-assisted analysisPrivacy was the deciding factor. The reviewer controlled the local agent from his phone through Telegram. He used it to sort a Steam library of more than 400 games and to analyze financial records and embargoed specs that he would not send to a cloud service. [1]
A local agent needs ongoing maintenance. A 7:30am briefing failed because macOS was asleep. Scheduled jobs start in fresh sessions, so each prompt has to contain everything the job needs. The agent also needed broad access, including a Steam API key that the reviewer revoked afterward. [1] [2]
Two things make the setup possible: a mixture-of-experts design that activates about 6B of the model's 125B parameters per token, and a lot of memory. One reviewer measured roughly 86–143 tokens per second of generation on the M5 Ultra. Tom's Hardware's top configuration as tested cost $12,299. [6] [8] [14]
one user's diary and one reviewer's benchmarks show sensitive files being handled entirely on one machine. Implication: for legal, finance and pre-release work, the cost moves from per-token fees to about $12,000 of hardware plus ongoing maintenance.
Executive brief
A Verge laptop reviewer with no AI background set up Nous Research's open-source Hermes Agent on a 256GB M5 Ultra Mac Studio. He ran Qwen3.8-Flash-Next, a 125-billion-parameter model, entirely on that machine and controlled it from his phone through Telegram (The Verge). Within days the setup was sorting a library of more than 400 Steam games and working through financial records he would never upload to a cloud service. His daily briefing job also broke several times. His conclusion: privacy, not raw capability, was what got him to use AI at all. A ~$12,000 machine makes that possible but doesn't make it magical.
What changed and event timeline
Hermes Agent ships
Nous Research released Hermes Agent as an MIT-licensed, self-hosted agent that works through messaging apps, writes its own skills and runs scheduled jobs ().
Injection flaw disclosed
CVE-2026-9353 (CVSS 7.3) covers hermes-agent versions up to 2026.4.23, in the Skills Guard component.
RTX Spark announced
Microsoft and Nvidia announced Windows PCs built for local agents, with up to 128GB of unified memory ().
Qwen3.8-Flash-Next released
Alibaba's Qwen team published open weights for a mixture-of-experts (MoE) model with 125B total parameters and 6B active per token. It previews the architecture planned for Qwen4 ().
M5 Ultra reviews land
Reviewers called the Mac Studio a "local AI powerhouse," and several tested Qwen3.8-Flash-Next and Hermes on it ().
Verge diary published
Di Benedetto describes his first local-agent workflows and where they failed ().
Capabilities and access
- Hardware: M5 Ultra Mac Studio with 256GB unified memory. Tom's Hardware's top configuration cost $12,299 as tested (Tom's Hardware).
- Model: Qwen3.8-Flash-Next, about 105GB on disk according to The Verge. The hosted version is sold as Qwen3.8-Flash (Codersera).
- Agent: Hermes is free when paired with local models. It can also connect to Ollama or to cloud providers (BestLLMFor).
Read the full section
- Hardware: M5 Ultra Mac Studio with 256GB unified memory. Tom's Hardware's top configuration cost $12,299 as tested (Tom's Hardware).
- Model: Qwen3.8-Flash-Next, about 105GB on disk according to The Verge. The hosted version is sold as Qwen3.8-Flash (Codersera). The model is also available through Ollama and other local runtimes.
- Agent: Hermes is free when paired with local models. It can also connect to Ollama or to cloud providers (BestLLMFor).
- Windows option: RTX Spark laptop preorders are open, with availability on Oct 16 (ASUS).
Technical analysis for researchers and developers
- Architecture: In three of every four layers, the model uses Gated DeltaNet, which compresses earlier context into a fixed-size state.
- Agent memory: Hermes keeps two small files, MEMORY.md and USER.md, and searches past sessions with SQLite full-text search (DataCamp).
- Scheduled jobs: Cron jobs start in fresh sessions, so each prompt needs to contain everything the job requires (19:18).
Read the full section
- Architecture: In three of every four layers, the model uses Gated DeltaNet, which compresses earlier context into a fixed-size state. The remaining layers use attention for exact retrieval. It also adds 51B parameters of N-gram embeddings and supports a 262K-token context window. It runs on vLLM, SGLang, llama.cpp and MLX (GitHub).
- Agent memory: Hermes keeps two small files, MEMORY.md and USER.md, and searches past sessions with SQLite full-text search (DataCamp). The video transcript describes these files as deliberately small and added to the prompt at the start of every session (11:13).
- Scheduled jobs: Cron jobs start in fresh sessions, so each prompt needs to contain everything the job requires (19:18). Separately, The Verge's briefing failed because macOS was asleep when the 7:30am job tried to run.
Claims and evidence
- Local speed: MacStories reports about 2,700 tokens/s for prompt processing (roughly 2.5× the M3 Ultra) and 86–143 tokens/s for generation on Qwen3.8-Flash-Next across 4- to 8-bit quantizations (MacStories).
- Training efficiency: Qwen says the model needs about one-ninth the training compute of Qwen3.7-Plus while performing better at coding and office tasks (GitHub).
- Task outcomes: The Steam sorting, data analysis and briefing results are one user's anecdotes, not a benchmark (The Verge).
Read the full section
- Local speed: MacStories reports about 2,700 tokens/s for prompt processing (roughly 2.5× the M3 Ultra) and 86–143 tokens/s for generation on Qwen3.8-Flash-Next across 4- to 8-bit quantizations (MacStories). This is a single reviewer's testing.
- Training efficiency: Qwen says the model needs about one-ninth the training compute of Qwen3.7-Plus while performing better at coding and office tasks (GitHub). This is the developer's own claim, and no independent replication appears in the reviewed sources.
- Task outcomes: The Steam sorting, data analysis and briefing results are one user's anecdotes, not a benchmark (The Verge).
Context and prior work
- Apple is selling its Macs on local AI performance and demonstrated a four-Mac Studio cluster worth nearly $50,000, according to The Verge.
- Nvidia and Microsoft are making the same pitch for Windows with RTX Spark (NVIDIA Newsroom).
- The video transcript presents Hermes as an "operating layer" that sits around a person's work, as opposed to a chatbot or a coding agent aimed at one project (01:22).
Read the full section
- Apple is selling its Macs on local AI performance and demonstrated a four-Mac Studio cluster worth nearly $50,000, according to The Verge.
- Nvidia and Microsoft are making the same pitch for Windows with RTX Spark (NVIDIA Newsroom).
- The video transcript presents Hermes as an "operating layer" that sits around a person's work, as opposed to a chatbot or a coding agent aimed at one project (01:22).
Limitations, safety and contested findings
- Local models don't fit every job: The video transcript says local models suit simple private tasks, but a weak one can waste time on messy multi-step workflows (09:11).
- Permissions: To act, the agent needed broad access, including a Steam API key, which the author revoked afterward.
- Security: The CVE-2026-9353 advisory says the vendor did not respond to the disclosure (Vulert).
Read the full section
- Local models don't fit every job: The video transcript says local models suit simple private tasks, but a weak one can waste time on messy multi-step workflows (09:11). It recommends top cloud models for reliable agent work (28:33). The Verge's experience partly bears this out: getting working benchmark-automation scripts was still "a work in progress."
- Permissions: To act, the agent needed broad access, including a Steam API key, which the author revoked afterward.
- Security: The CVE-2026-9353 advisory says the vendor did not respond to the disclosure (Vulert).
- Human approval: The video transcript recommends a human checkpoint before external or irreversible actions such as sending email (17:16).
Business and practitioner implications
- Privacy can be the deciding factor: Embargoed specs and financial data were the tasks the author would only hand to a local model.
- Hardware is the cost: A capable local stack currently means 128–256GB of unified memory and roughly $12k for the Mac configuration tested.
- Operations matter: Treat local agents like software that needs maintenance.
Read the full section
- Privacy can be the deciding factor: Embargoed specs and financial data were the tasks the author would only hand to a local model. This matters for legal, finance and pre-release work.
- Hardware is the cost: A capable local stack currently means 128–256GB of unified memory and roughly $12k for the Mac configuration tested. There are no per-token fees after that.
- Operations matter: Treat local agents like software that needs maintenance. Keep scheduled jobs self-contained, keep machines awake, give out narrow and revocable credentials, patch Hermes, and require human approval for actions that leave the machine.
- Start small: The video transcript advises adding one model, one interface and one workflow at a time, and favoring jobs that only alert when something important changes (25:09).
The source trail.
Sources (16)
Learning to use local AI is exciting, overwhelming, and frustrating
Article text retrieved; extracted text may omit tables or interactive elements. A transcript of the video linked from this page is supplied (linked video auto captions; https://www.youtube.com/watch?v=Ta2wg6xPaY4); automatic text may contain errors.
theverge.comLinked video: Learning to use local AI is exciting, overwhelming, and frustrating
linked video auto captions
www.youtube.com