devPad

We should collab! There is an open dataset available on the site. If you have data to share I can integrate it. If you have runs you want to submit, download the latest release of anybis-oss from GitHub and submit as many benchmarks as you want, all the data is open source and I recompile the report periodically github.com/uncSoft/anubis…

@devPad_JT @dealignai oh this is great! can you include my reports there?

ONE OF THE MOST INSIGHTFUL FASTEST READS FOR ALL MAC LLM USERS. In 2026 with the rise of the M5 Max, the MLX community is becoming flooded with models and choices, where I completely understand if someone becomes overwhelmed due to the large amount of choices. x.com/allenwlee/stat… Allen excellently reviews and shows the current meta of utilizing LLM's in the Mac environment, going out of his way to review topics that so many of us are iffy about when it comes to M chips; TTFT and coherency. #macbook #m4max #m5max #mlx #llm #qwen3

The Results Are in! Hey all - I've been working on Anubis OSS github.com/uncSoft/anubis…, an open source macOS app for benchmarking local LLM inference on Apple Silicon. It tracks tok/s, TTFT, power draw, GPU/CPU utilization, memory pressure - basically everything happening on your Mac while a model runs. Even has a built in standalone performance monitor and light system benchmarking. The repo just broke 100+ stars which is amazing for my first open source project. We've collected over 150 community benchmark runs across 36 users, 85 models, and 8 Apple Silicon chips so far. Finally got around to putting together an analysis of the results. Some highlights: M4 Mac mini is the efficiency king - ~8W system power, 5.35 tok/W. Punches way above its weight class. MoE models are the move on Mac - 120B parameter MoE models running at 70+ tok/s on M4 Max because only ~10B params activate per token. If you're not running MoE yet, you're leaving performance on the table. Backend matters more than you'd think - MLX consistently beats Ollama by 5-10% on small models on the same hardware. Same model, same Mac, different numbers. The MacBook Neo (A18 Pro) can actually do it - 50 tok/s on 1B, 23 tok/s on 3B. Don't try anything bigger than 7B though. There's a lot more in the full report - throughput charts per chip, memory bandwidth correlations, TTFT analysis, a top-15 leaderboard, big model (100B+) breakdowns, etc. 👉 Full Benchmark Analysis Report devpadapp.com/anubis_bench_a… 📊 Live Leaderboard devpadapp.com/leaderboard.ht… - upload your own runs ⬇️ Download Anubis OSS v2.9.0 github.com/uncSoft/anubis… - dev cert signed, auto-updates via Sparkle The app is free/open source (GPL-3.0), native SwiftUI, macOS 15+. Would love more data points - especially from M1/M2/M3 Ultra owners and anyone running weird model configs. The more runs we get the better this gets. I really love programming this app - and a new refresh is coming soon (way more users than I ever thought!) GitHub github.com/uncSoft/anubis… Main Site devpadapp.com/anubis-oss.html

@minchoi This is what happens when you use Claude to build Claude

RIP OpenClaw 💀 Claude now has > Voice mode > Agent Teams > 38+ Connectors > Cowork Projects > Scheduled tasks > Plugin Marketplace > Persistent memory > 1M Context window > Dispatch for remote control > Channels for Telegram & Discord > Claude can use computer to run apps💀

@alexocheema If someone can test a cluster on Anubis, that would be so helpful, check the report if submitted runs, pretty cool devpadapp.com/anubis_bench_a…

The new M5 Pro/Max MacBooks have 3 Thunderbolt 5 ports, enabling you to create RDMA clusters with up to 4 MacBooks. The latency with RDMA over Thunderbolt is single digit microseconds, fast enough for tensor parallelism with close to linear scaling.

@lucatac0 Submit the scores to our open dataset! devpadapp.com/leaderboard.ht… github.com/uncSoft/anubis…

first test on the new M5 Max MBP Qwen 3.5 27B running locally at 32 tok/s via mlx

@dealignai Submit some runs to the Anubis leaderboards, tons of good data there . 101 stars on GitHub! devpadapp.com/leaderboard.ht…

For MacBook 128gb owners - MiniMax m2.5 at 60gb in JANG format should be your base model. While 120b models are great, the kinds of understanding that a 230b model has is a noticeable difference. Running at a smooth 45+ token/s on M4Max/M5Max, this has been my go to LLM for alot of work. #macbook #m5max huggingface.co/JANGQ-AI/MiniM…

@2BGood4Now @ZenLightningBug @BillMelugin_ @grok Who do you think pays to fly them aound now, you retard

@ZenLightningBug @BillMelugin_ @grok Many did under O’Biden, always at taxpayer expense.

BREAKING: President Trump says if Democrats don’t immediately agree to a deal on DHS funding, he will place ICE agents at airports to conduct security, “including the immediate arrest of all Illegal Immigrants who have come into our Country”.

@danburgh1 @minugirl49268 @BillMelugin_ Perhaps Republicans should pass the TSA spending bill they rejected

@minugirl49268 @BillMelugin_ It has to get done. Perhaps the Military could handle it. But, it is situations like this that Executive Orders are Necessary and Appropriate. It is Criminal for Congress to leave all these Ports of Entry unattended over Political disagreements. The POTUS must handle it.

@MaryMMarti35582 @Sirius420Nova @BillMelugin_ Deporting without a hearing is not ensuring security.

@Sirius420Nova @BillMelugin_ "Both ICE and TSA operate under the Department of Homeland Security (DHS), which is part of the executive branch. The President generally has authority to deploy federal law enforcement personnel, such as ICE agents, within the U.S. to ensure security."

@BillMelugin_ They aren't getting paid either 😂😂😂🤡🤡🤡

@lamps_apple @stoolpresidente You're saying this from the only major city in the US MORE liberal than NYC lmao stfu

Florida has NO state income tax and sunshine... and over 125,000 New Yorkers moved there, taking nearly $14 BILLION in income with them. Meanwhile, "Begging Kathy" Hochul is holding a PRESS CONFERENCE asking them to come back!! Her radical socialist NYC Mayor Zohran "Hamas" Mamdani wants to push the combined city-state tax rate to 16.78%... the HIGHEST IN THE NATION... and add federal obligations bringing the total burden to nearly 54%. President Trump WELCOMES them all to FLORIDA and TEXAS where they are RESPECTED. That's called LEADERSHIP!

The unbelievable arrogance and hypocrisy of begging millionaires to return to New York while the new Mayor simultaneously says he despises millionaires and supports communism. I wonder why people are flocking to Florida?

@therealCSmith57 @Baldwin1970 @NYMag Just make shit up

@Baldwin1970 @NYMag People are leaving the state at such a high rate the governor asked wealthy people to please return so they could fund social programs

Mayor Zohran Mamdani recently got the political equivalent of what baseball players call a brushback pitch — a fastball deliberately thrown dangerously close to a batter’s head in order to intimidate the player, who must flinch or duck to avoid a devastating injury. The mayor is getting municipal chin music from the major bond-rating agencies: Moody’s formally changed its outlook on the city’s finances from “stable” to “negative,” and S&P Global Ratings opined that Mamdani’s budget plan will “make it difficult to sustain budgetary balance beyond fiscal years 2026 and 2027.” The negative outlook from the agencies is a warning, writes columnist Errol Louis. The next step could be a downgrade of the city’s bond rating, which would raise the cost of borrowing money for routine city operations. Mamdani maintains that the decision to revise the outlook is premature, pointing out that the city’s overall credit rating remains strong and has not been downgraded. But the message from Wall Street seemed crystal clear: Unless Mamdani adopts a more fiscally conservative approach, we will punish City Hall in the markets. Read Louis’s full column: nymag.visitlink.me/H057m5

@danpacary @supziez Coo!l how can others submit their runs through a multi backend supported benchmarking app

@devPad_JT @supziez alright done bench.rustane.org

If you're on mac silicon... M-series Here is what you can do (projected) remember this is training validation x.com/danpacary/stat…

@Suryanshti777 Get some more ollama scores on the leaderboards! devpadapp.com/leaderboard.ht… github.com/uncSoft/anubis…

🚨 BREAKING: your laptop is officially powerful enough to run ChatGPT. no cloud. no API. no limits. this is llama.cpp — and it’s insane: → run LLMs locally in pure C/C++ → works on CPU, GPU, even low-end machines → 1-bit to 8-bit quantization (crazy efficiency) → supports Apple Silicon, NVIDIA, AMD, even RISC-V → OpenAI-compatible API server in ONE command → pull models directly from Hugging Face people are still paying per API call… while others are running their own AI stack locally. this changes everything: privacy ✔️ cost ✔️ full control ✔️ this isn’t just another repo. this is the foundation of local AI. repo in comments 👇

@danpacary @supziez Sure it's super easy, any idiot can do it! Open source it and share it when you're done

@devPad_JT @supziez Or I could just make a leaderboard

@danpacary @supziez devpadapp.com/leaderboard.ht… github.com/uncSoft/anubis…

@devPad_JT @supziez Yeah, I should totally start a leaderboard

@supziez @danpacary Submit those scores to the leaderboards through Anubis!

@danpacary M5 MacBook Air 32GB 10C/10G

@BrianRoemmele if you're interested in submitting some runs to the LLM leaderboards through Anubis, we just added oMLX support - would love to get the devs some mac hardware and model specific data github.com/uncSoft/anubis… Hundreds of runs uploaded so far devpadapp.com/leaderboard.ht…

Testing this now. Quite useful on my laptop, omIx LLM inference server with continuous batching & SSD caching for Apple Silicon - managed from the macOS menu bar github.com/jundot/omlx