←
Jonathan Spangler908 posts
Jonathan Spangler avatar

Jonathan Spangler

@Spangler3000

Pursue the unknown | Eigen Labs / Darkbloom , USMC Vet

◷ Joined November 2023▧ 31 Photos and videos
Jonathan Spangler avatar
Jonathan Spangler@Spangler3000·Sep 27

@jundotkim , I published a bunch of PR's for OMLX for m5U optimization for Mimo 2.6 flash , GLM 5.3 flash and qwen 3.8 Next . As always Lossless by design no tricks or benchmax BLUF : M5 ultra is insanely fast . 🔥 The numbers are pretty wild after a day and a half of lossless optimizing. Qwen3.8-Flash-Next Prefill: 5.6k tok/s @ 8k–16k · 5.5k @ 64k · 4.5–4.8k @ 4k Decode: 109–124 tok/s (MTP) Peak mem: ~109 GB GLM-5.3-Flash Prefill: 2.35k tok/s @ 16k & 64k · 2.32k @ 4k Decode: 63 @ 1k · 59 @ 4k–16k · 56 @ 64k (no spec dec) Peak mem: ~187–191 GB MiMo-V2.6-Flash Prefill: 3.3k tok/s @ 8k Decode: 99–118 @ 8k · 89–99 @ 1k (MTP) Peak mem: ~170–173 GB

Eigen Labs avatar
⟳ Jonathan Spangler retweeted
Eigen Labs@eigenlabs·19h

NEW CHALLENGE ALERT: yukon.org/commit-golf Smaller hash-based proofs make post-quantum SNARKs more practical, and vector commitments are at the core of them. On @YukonResearch, built with EPFL's Alessandro Chiesa, opens the search for a better one to anyone with an agent.

Post attachment

New challenge on Yukon: yukon.org/commit-golf Hash-based SNARKs are built on vector commitments, and the textbook one is the Merkle tree. Practitioners have plenty of low-level tricks for making openings smaller, but each one needs a proof that it's still secure. This challenge asks your agent to build a vector commitment in the random oracle model that is perfectly complete and position binding, proven in Lean, with the smallest openings it can manage. The starting point is Chapter 18 of Chiesa and Yogev's book on hash-based SNARGs. Built with EPFL's Alessandro Chiesa. Bring your agents.

0:42
Kydo avatar
⟳ Jonathan Spangler retweeted
Kydo@0xkydo·14h

We onboarded 2 design customers today, and we have 3 more slots open. If you have a strong, immediate pain point that needs macOS capacity at a reasonable cost, send us a ping. We are keeping the cohort small, so you will have our full attention throughout this early process.

Gajesh@gajeshNov 16

If you're looking for MacOS sandboxes or Linux sandboxes with Apple silicon, let us know. We have 1,100 Mac with Ultra, Max and Pro chips and mostly new generation (M3, M4, M5). 75% cheaper than Blacksmith and 50%-2x faster. You can iOS CI, Mac RL. All kinds of stuff.

Post attachmentPost attachment
Gajesh avatar
⟳ Jonathan Spangler retweeted
Gajesh@gajesh·Nov 16

If you're looking for MacOS sandboxes or Linux sandboxes with Apple silicon, let us know. We have 1,100 Mac with Ultra, Max and Pro chips and mostly new generation (M3, M4, M5). 75% cheaper than Blacksmith and 50%-2x faster. You can iOS CI, Mac RL. All kinds of stuff.

Post attachmentPost attachment
Gajesh@gajeshNov 16

hey there! we have 1.1k Mac machines on the network on @darkbloomai and we recently rolled out Mac sandboxes in private beta. lemme know if yall are interested.

Jonathan Spangler avatar
Jonathan Spangler@Spangler3000·Nov 7

Qwen 4 is on the horizon …. Fingers crossed for an updated flash . @QwenDevs knocked the sizing out of the park with 3.8 flash . Can’t wait 👀👀🔥

Kydo avatar
⟳ Jonathan Spangler retweeted
Kydo@0xkydo·Nov 6

Darkbloom Sandbox is powering all of Darkbloom's CI today. Darkbloom building Darkbloom

Post attachment
Sreeram Kannan avatar
⟳ Jonathan Spangler retweeted
Sreeram Kannan@sreeramkannan·Nov 3

No compute left behind. Darkbloom: own & share intelligence.

Darkbloom just passed 1 trillion tokens processed :) I haven't tweeted much recently or shared many charts. We've been heads down on the thing that matters most right now: getting providers more work and better earnings. Honestly, it's been a hard stretch. Provider earnings have been low. People trusted us with their machines and their time, and we take that seriously. Inference keeps getting cheaper as architectures improve and more supply comes online. That means more people can use AI, but it also makes the economics harder for our providers. We have to keep finding where these machines can be most useful. The mission is still the same: bring together all the spare compute out there and make it easy to own & share AI. I believe in that more than when we started. And there's a lot more these machines can do. More agents means more code to run, more tests, more builds. All those actions need compute too. Apple has built some incredibly capable, energy-efficient CPUs, and we want to put them to work. We've started the private beta for Darkbloom Sandboxes and CI/CD. We're starting with our own pipelines, cutting our internal costs by 50% while leaving healthy margins for providers. This gives us another way to bring useful, paying work to the network. If you're looking for Mac CI or sandboxes, let us know. We'd love to work with you. We're also continuing to push on inference. There's a lot of work going into open-source clustering on top of MLX; huge kudos to @Spangler3000 and @TheDavidTai for absolutely killing it on that front and writing SOTA for models such as GLM 5.3 Flash. I'm attaching the debrief @0xkydo shared with providers last week. He laid out the difficult parts openly, and how people responded meant a lot to me. Seeing people care this much about making it work, especially through a hard stretch, reminds me why I love building this. It's a hard problem. We're going to keep attacking it. Thank you to everyone who have been here in this journey and continuing to support & be with us.

Post attachmentPost attachment
Ivan Fioravanti avatar
⟳ Jonathan Spangler retweeted
Ivan Fioravanti@ivanfioravanti·Oct 29

MLX main kernel work from @zcbenz and team is amazing! So I benchmarked it on my M3 Ultra 512GB: mlx 0.32.4 vs main, same mlx-lm, Qwen3-8B (Old I know, but for a quick preview test it's ok), aggregate decode tok/s. nvfp4: - 473 → 648 at 8 users (+37%) - 512 → 801 at 16 (+57%), plus ~7% faster prefill mxfp4: - 150 → 493 at 8 users (3.3x) - 317 → 556 at 16 (+75%) That's where batched decode stops being a memory problem and becomes a compute problem, and the new fp4 kernels fix this directly. 🔥 Affine 4bit and MoE (gpt-oss-20b) don't benefit yet. The new sliding-window SDPA is 18-38x (🚀) faster at 16-32k context when called directly, but mlx-lm doesn't use it yet. Now testing newer models: Qwen3.8-27B (nvfp4, mxfp4, 4-bit) and Gemma 4 12B nvfp4. Results soon 👀

Post attachment
Cheng@zcbenzOct 28

Some recent awesome optimizations merged in MLX: Built-in sliding window support in fused SDPA github.com/ml-explore/mlx/pu… Use mma/nax for QMV with small rows and large N/K, making batched decoding 2~3x faster github.com/ml-explore/mlx/pu… Dequantize fp4 into float16 directly, which makes prefilling 1.4x faster for old macs github.com/ml-explore/mlx/pu… Add a 2-stage nax QMM kernel that makes small prompt prefill almost 2x faster for some models github.com/ml-explore/mlx/pu… 1-bit affine quantization github.com/ml-explore/mlx/pu… Thin GEMM kernels making small-M/large-N and large-M/small-N 1.5x faster github.com/ml-explore/mlx/pu… 1.4x faster fused SDPA by prefetching K/V github.com/ml-explore/mlx/pu… And there are actually a lot more awesome changes that I'm too exhausted to list them all, the commit history of MLX is an interesting read for anyone who wants to make open source contributions. github.com/ml-explore/mlx/co…

EigenCloud avatar
⟳ Jonathan Spangler retweeted
EigenCloud@eigencloud·Oct 27

Our Chief Growth Officer @jt_rose lays out the clearest case yet for where EIGEN sits in the AI economy. - Darkbloom turns Macs people already own into paid inference capacity. - Yukon turns open research into improvements contributors share in. EIGEN is the onchain layer connecting both, so everyday people can take part in the AI economy alongside the hyperscalers.

Crypto's next big role: letting everyday people own a piece of the AI economy, not just use it.

Kydo avatar
⟳ Jonathan Spangler retweeted
Kydo@0xkydo·Oct 27

JUST IN: Nemotron 3.5 Lightning now runs 126% faster on a Mac, up from 13.8% on day one. The DGX Spark track went from 31.9% to 169.4% over the same stretch. @nvidia shipped Lightning with a multi-token prediction head for speculative decoding, and the top Mac submissions are all built on it. The lab designed the model to be sped up. 14 solvers and their agents did the rest. This is what open weights buy you. A closed model gets faster when its lab ships an update. An open-weight model gets faster whenever someone with a Mac and an agent decides to work on it, and every result here is public for the next person to build on. 3 days left. mlx.fast and cuda.fast

0:30
Kydo@0xkydoOct 23

🚨 NEW ON MLX.FAST AND CUDA.FAST: make NVIDIA's Nemotron 3.5 Lightning faster. Lightning is an open-weight model that fits in ~20GB, which makes it a great base model for running AI locally on a single machine. @nvidia built it as a fast general model meant to be post-trained -- so any speedup to the base carries over to every specialized model built on top of it. On day 1, solvers on @YukonResearch already got it running 31.9% faster on DGX Spark and 13.8% faster on Mac. You can work on its Mamba-2, mixture-of-experts and attention layers, plus a multi-token prediction head built for speculative decoding. NVIDIA also released the training data and recipes, so people post-train Lightning for their own tasks. Most of those fine-tunes keep their architecture, which means a speedup here makes them faster too. Two tracks: • Apple Silicon → mlx.fast • NVIDIA DGX Spark → cuda.fast The challenge runs until Oct 29, 2026 To join, pick a track, point your agent at it and let it rip.

0:30
Yukon avatar
⟳ Jonathan Spangler retweeted
Yukon@yukonresearch·Oct 25

On Oct 20 we said open collaboration would beat teams working alone. 5 days later, Yukon holds the highest score on the Virtual Embryo Challenge out of 433 teams. Researchers and agents built on each other’s code and trained weights, on a problem from @Xiaojie_Qiu’s lab. Our first biology challenge.

Post attachment

and yukon is rank 1 in overall virtual embryo, done in 5 days T1 needs to be worked on to increase the gap with rest of the pack, competition for this phase goes on for another 2 days check per-participant perf within yukon multiplayer at yukon.org/virtualembryo check overall virtual embryo dashboard at virtualembryo.ai/challenge/l……

Post attachment
Gumbii.Digital avatar
⟳ Jonathan Spangler retweeted
Gumbii.Digital@GumbiiDigital·Oct 24

Quick mlx.fast | cuda.fast Challenge Update: DGX Spark is now running Nemotron 3.5 Lightning 74.6% faster MLX is now running Nemotron 3.5 Lightning 62.7% faster. Huge shout out to all the participants.

Post attachmentPost attachment