@ddalcu

I build OpenSource things, local LLM engines & UI’s. Created https://xitter.cf/t.co/qxuPXH75o1 for Desktop and iOS - https://xitter.cf/t.co/Mdoym1KQbp

github.com/ddalcu
Joined June 2010
  • Tweets1.1K
  • Following931
  • Followers2.8K
289Media
github.com/ddalcu/mlx-serve/… Many new models DeepSeek-V4.1-Flash is much faster Chat on the Neural Engine Meeting notes from the menu bar Speech to text transcriptions api RDMA Images, video, music and 3D up to twice as fast on M5 Macs Frontier models on a 128 GB Mac via Streaming
15
10
4
115
7,244
David Dalcu retweeted
BennyDaBall
@BennyDaBall_OG17h
There are so many DiT text-to-image models out right now I had to compare them all! Seven open models, same prompt, same seed, stock ComfyUI INT8 on my 5090. 1024x1024. My numbers (seconds per image): - Z-Image Turbo: 1.4 s - Qwen Image 2.1 Turbo: 1.8 s - FLUX 3 Klein: 1.9 s - Krea 2 Turbo: 2.5 s - Qwen Image 2.1: 4.2 s - MOSS-Image SFT: 5.8 s - FIBO 2 Dev: 17.7 s Honest take: every one of them nailed the scene. On this seed all seven spelled both signs right. The gap now is style and speed. Qwen 2.1 Turbo kept up with base Qwen at 2.3x the speed, and FIBO 2 is the slow one but has the most natural light and nailed the composition. MOSS-Image is a bit over saturated but has the most intrigue. All local. Which one would you choose?
10
8
4
141
18,032
David Dalcu retweeted
UNALIENABLE RIGHTS
@F1sT12h
#movies alternate universe takeover. FINAL #starwars
16
51
4
313
21,896
Currently working on 2 things. 1. Full re-write of MLX-Serve to be composable, support plugins/engines. 'mlx-serve plugin install deepseekV10' 2. Full re-write of the Swift code, can be powered by any endpoint(s). Both CLI + GUI, cross-platform.
7
2
1
91
3,087
Most downloaded MLX text models in October. (HF Data)
8
6
1
90
8,739
Btw new MLX-Serve is out, it supports FastH3 Omni, @Recursive_SI Flint, and much more. github.com/ddalcu/mlx-serve/…

Releases · ddalcu/mlx-serve

Native LLM inference server for Apple Silicon. OpenAI + Anthropic API compatible. No Python. Zig backend, Swift frontend macOS app with chat, music, voice, video generation. - ddalcu/mlx-serve

github.com
Hao AI Lab
@haoailabOct 20
As promised, FastVideo open-sources @MiniMax_AI FastH3 Omni, training code included! 🚀 - Real-time gen @ 12x faster on NVIDIA Blackwell - NVFP4/FP8 for Spark + RTX! - ComfyUI day-0 Collab with @NVIDIAAI , Nuva Lab @nuvalab and Reactor @reactorworld Open source must win! ♥️
6
4
2
56
5,381
David Dalcu retweeted
Tencent Hy
@TencentHunyuanOct 23
🚀 Introducing Hy-Nomas: An Open Source Free-Form Interactive Video World Model Key Features: * Free-form interaction with text prompts and camera trajectories * Long-horizon generation with world memory * Lightweight LoRA adaptation for specialized domains such as robotics * Optimized for Inference on Consumer GPUs Test it and drop feedback anytime！ 🔗 Website: hy-nomas.hy.tencent.com/ 💻 GitHub:github.com/Tencent-Hunyuan/H… 🤗 Hugging Face: huggingface.co/tencent/Hy-No… 🎥 Demo: hy-nomas.hy.tencent.com/demo #Hy#WorldModels #VideoGeneration #GenerativeAI
36
121
50
900
59,994
David Dalcu retweeted
JacksonS
@shrmodelxOct 23
Lmao😂#NVIDIA
144
784
131
6,692
912,746
Both of these UI's are native code. One is Swift ~80MB only works on Mac One is Zig ~10MB works on Mac/Linux/Windows
29
17
3
462
38,064
David Dalcu retweeted
anabology
@anabologyOct 22
Universal High Income The most likely outcome
550
782
383
5,876
976,651
David Dalcu retweeted
logicgate
@thelogicalgateOct 21
Abliterated (for cybersecurity) Sage-A1.5 hitting peaks of ~140 TPS running on @ddalcu's mlx-serve. Tested with writing php2shell exploit chain, as well as setting up its own test container to run it against. After completion it was still averaging over 100 tps decode and 2k prefill. Will post the hf link for the abliterated model in the comments when I get some more tests done.
4
4
1
6
562
Here is MiniMax H3 vs FastH3 running on an M5Ultra 80c Its a bit faster... 832×480, 243 frames
3
1
2
29
2,829
This is crazy, all local generated 🤯 MLX-Serve + all the models it supports can make some really cool games ! Deepseek 4.1 @deepseek_ai Moss image gen @Open_MOSS Trellis for 3D @MSFTResearch Yue2 for music @abc43992899 Stability AI SFX @StabilityAI
3
1
34
1,379
Maybe I can fine-tune it a bit, and push it somewhere so we can all play together. GTAI - Grand Theft Artificial Intelligence.
3
371
1
6
597
Picture -> 3D Rendered GLB models don't get enough love. Claude / ChatGPT / Deepseek 4.1 are good at games, but takes a very long time to create a good looking 3D model. With an Image model + Trellis2 and Pixal3D it takes a few seconds. Coming soon...
1
1
11
1,363
1
312
Krea 2 Turbo vs MOSS... I think these might be the best image models right now. Which one do you think looks better ?
3
12
1,522
Ivan FTW ! Thanks, merging soon ! Some more performance additions from my poor 5060ti and pushing the merge button. MLX, DGXSpark, 5060.. Linux... what is going on 🤯 ... oh Windows too... 🤣 github.com/ddalcu/mlx-serve/…

perf(cuda): Flash Next on a DGX Spark: prefill 2.5x, decode 1.8x with or without MTP by ivanfiora...

NVIDIA's Qwen3.8-Flash-Next-FP4-Local on a DGX Spark (GB10): prefill 2.5x and decode 1.8x with or without MTP, with outputs bit-identical to the paths they replace or no further from an f32...

github.com
Ivan Fioravanti
@ivanfioravantiOct 19
In this right moment I have agents working in parallel on different optimizations on same engine, on various machines: - mlx-serve on 3 x DGX Sparks - ds4 on 2 x M3 Ultra I never tried this, but it's fast! The bottleneck is the environment now, the agent and model are fast!!!
1
1
13
5,028
David Dalcu retweeted
Anemll
@anemllOct 17
Confirmed 24K context on iPhone with both K/V on int8. Delta net stays FP16. Decoding also peaks at 70 tok/s with easy to predict context. Apple Neural Engine on A20 is fast 🚀
Anemll
@anemllOct 17
Qwen 3.8- 27B on iPhone 18 Apple Neural Engine. Added draft model , dflash-2 at ~ 2.8 bpw ( mixed 2-4 vector quantization with small drop in acceptance rate) Base is @PrismML Bonsai-2, 1.585 bpw requantized to 1.5 bpw ( VQ 6 bits per 4 weights) Decoding speed is 15-60 tps. Fixed context size 16K. We might push it to 24K. 6GB to 7GB iOS App ceiling is the main limiting factor. I like this as PoC . Any suggestions for actual iPhone App on ANE?
5
10
2
61
3,334
Load more