John Records
24 posts
@sudoingX Or a used m1 Max MacBook pro with 64 GB for around $1,300 with a 1-year warranty refurbished, on eBay
English
this is the right read, and it's the box i'd point a newcomer at. the 64gb framework is the cleanest on-ramp into moe local ai there is.
same chip as my 128gb, so identical speed, not a cut down part. 64gb holds the 35b-class moe models at full quant with room for context, which covers most of what you'd actually run day to day.
and it's a clean integrated box, no discrete gpu to wire in, no psu to cable, you drop in storage and you're running models.
$1,959 for the chip, ~$2,253 configured like this one with storage. a complete, quiet ai machine that punches above its price. the 128gb is for people loading the giants. for getting started with moe, the 64gb is genuinely the move.
here is the link:
frame.work/products/deskt…
exitLQ@CdeBurner
@sudoingX with current prices i think the 64gb version is the sweetspot for me
English
@Prince_Canuma @GoogleDeepMind A google engineer posted that apple hardware will not provide the amazing speed boost that separate gpu provides. What's your experience? Thanks for your great work.
English
mlx-vlm v0.6.3 is here 🚀
Day-0 support for TWO new models from our partners we work closely with:
🔥 @GoogleDeepMind DiffusionGemma — a genuinely new architecture. Instead of token-by-token, it generates 256-token blocks in parallel with bi-directional attention and iteratively self-corrects the whole block, image-generator style. 26B MoE, only 3.8B active, fits in 18GB quantized. Day-0 MLX support via our Google DeepMind partnership, with long-context prefill tuned and ready.
🔥 @cohere's North Mini Code 1.0 — a 30B MoE with just 3B active, running ~66 tok/s in BF16 before any compression. Day-0 on MLX thanks to our close collaboration with the Cohere team.
Get started today — install from source:
> uv pip install -U mlx-vlm
Then serve the model and point your favorite agent at it (pi, opencode, hermes, etc.):
uv run mlx_vlm.server --model MODEL-REPO
Model collection 👇🏽
Google Gemma@googlegemma
Meet DiffusionGemma! An experimental open model that explores a fast approach to text generation, released under an Apache 2.0 license. Moving beyond sequential, token-by-token processes to generate entire blocks of text simultaneously. Here’s what’s new with DiffusionGemma: 👇
English
One thing to keep in mind as you try DiffusionGemma – you'll want a dedicated accelerator (GPU or TPU) to see real speedups.
In particular, we love our MacOs AI developers, but this model may not be best for you!
Google Gemma@googlegemma
Meet DiffusionGemma! An experimental open model that explores a fast approach to text generation, released under an Apache 2.0 license. Moving beyond sequential, token-by-token processes to generate entire blocks of text simultaneously. Here’s what’s new with DiffusionGemma: 👇
English
English
This is my first-ever open-source project.
github.com/alxgntv/OpenWa…
This app turns your @Apple Watch into an AI-powered, wrist-first interface for @openclaw agents. fyi @steipete
It hasn't been published to the @AppStore yet, but I'll do that soon. For now, it can be installed via @Xcode and paired with your watch manually.
It supports a ton of features:
1. Displaying your main and sub-agents
2. Chat history for each session with each agent
3. Haptic feedback
4. Welcome messages
5. Voice responses
6. Greeting messages
7. Works even when your iPhone is locked
For now, it only works on Apple Watch, but I'll be expanding it step by step for other devices.
Built entirely with @cursor_ai.
Looking for feedback.
English
@hermes_updates Downloaded, and it wants to install Hermes. But Hermes already is on the computer. I paused the installation since I’ve got it working nicely and don’t want to mess it up.
English
@HermesAgentTips refurb M1 Max MacBook Pro 64 gb. Much cheaper than equivalent Mac Studio. Benchmarks with different MLX inference engines bright-lotus-8q5y.here.now
English
im trying to gather some knowledge of off the local LLM folks on X...
What's an affordable solution for someone that wants to get into learning the basics of the local LLM scene and dont have the cash to drop on expensive hardware
drop some hardware spec/models lets see what yall suggest
English
Which inference engine for local Mac Hermes? rapid-mlx benchmarked best on M1 Max 64 gb. YMMV. bright-lotus-8q5y.here.now
English
An immersive mythological fantasy that blends ancient wisdom, spirituality & modern-day realities into a captivating narrative. Set across different parts of India during the 80s & 90s, the book follows 3 people whose lives are unknowingly connected to a much larger cosmic design
English
@signalgaining Excellent, looking forward to more on your Jetson cases and WendyOS.
English
The NVIDIA Jetson Orin Nano is perfect for developers, but it comes naked, without a hard drive or an operating system. We're changing that, ready to go with WendyOS installed, just plug and play!
English
@HermesAgentTips deepseek-v4-flash which is free at the moment on nous
English
whats your favorite model to use with Hermes agent?
English
@jinyuhou0 @vishalm4341 I’ve looked for the link to the models, no luck. I’m eager to see it! Please consider posting it conspicuously, perhaps in its own tweet. Thanks, mate!
English
@vishalm4341 Yes! Everything is in the last post of the original thread (3/3) — code and models are all open.
English
On popular benchmarks, our 30B model matches systems 20-30x its size (gpt-5.4-xhigh, DeepSeek-V3.2, Kimi-K2.5), while using up to 95% fewer reasoning tokens than comparable 30/32B agentic LLMs.
The trick: don't just reason less, reason about the right things. A learned configurator decides when to simulate, how far ahead, and when to skip planning entirely.
Efficient reasoning is an allocation problem, not a compression problem.
Model and code are openly available.
Mingkai Deng@mdeng34
Frontier LLMs are converging on efficient, adaptive reasoning. Opus 4.7 lets the model decide how deeply to reason. GPT-5.5 achieves strong results with fewer reasoning tokens. We study a related but more structural question: what 𝗸𝗶𝗻𝗱 𝗼𝗳 𝗿𝗲𝗮𝘀𝗼𝗻𝗶𝗻𝗴 should we adapt? Last year in SiRA (upper figure), we showed that simulative reasoning (System II), which uses a 𝘄𝗼𝗿𝗹𝗱 𝗺𝗼𝗱𝗲𝗹 to evaluate consequences of actions, yields up to 124% improvement over reactive baselines (System I), and that strong reasoning models (o1, o3-mini) fail as planners without this structure. In our new paper SR²AM (lower figure), we add a learned 𝗰𝗼𝗻𝗳𝗶𝗴𝘂𝗿𝗮𝘁𝗼𝗿 (System III) that self-regulates when to simulate, how far ahead, and when to skip planning entirely. Efficient reasoning is not just shorter reasoning: it is better allocation of simulation.
English
We are making our discount permanent! 🎉
Enjoy building with DeepSeek-V4-Pro and bring your innovative ideas to life! 🚀
DeepSeek@deepseek_ai
The DeepSeek-V4-Pro discount has been extended until May 31, 2026, 15:59 UTC!
English
@rubengarciajr @rabbit_hmi Agreed and working that. Rabbit staff say they expect to release Rabbit update with Hermes availability.
English
@rabbit_hmi You should install Hermes on this thing and make it worth it
English
@HermesAgentTips I haven’t seen any studios at that price. But ebay has Mac laptops with those specs for about that price
English
everyone's buying $5,000 GPUs to run local LLMs
meanwhile a used Mac Studio M1 Max 64GB is doing 60+ tok/s on Qwen3 35b for $1,500
silent. cool. holds resale value.
English
@JohnRecords Mirage.elipnik.com has pricing. Still not in stone yet but good indicator
English
Currently using Mirage while my laptop charges at my desk and I know at this point I shouldn’t be surprised but it works so well. Super smooth and pixel perfect. Looks and feels like macOS is running on my iPad. No other app feels this smooth and uses touch this well.
English