
founded @ixopay. now investing, advising and in a love-hate relationship with local AI. let's see where this goes 🙃
Vienna
Joined July 2013
- Tweets2.9K
- Following236
- Followers153
- Likes1.4K
Rene Siegl retweeted
🐨 somora v2026.1116.11
🛠️ coding-agent teams in tmux
🔔 agents wake when coding agents need them
🖥️ a start menu behind the koala
💬 answered questions stay in chat
notes 👇
github.com/thenaxon/somora_a…
🚀
🐨 somora v2026.1111.1
🔧 leaner builder rules
🛡️ security fixes from a codebase review
⚙️ per-model builder settings
🔔 cron ranges now work
📦 Codex 0.167 + Grok 1.0.80
notes 👇
github.com/thenaxon/somora_a…
1
38
Rene Siegl retweeted
🐨 somora_realtimevoice is open source
Voice calls on your own models: your speech recognition, your LLM, your voice. Clone one from a short recording.
🔌 OpenAI Realtime-compatible
🗣️ barge-in, room noise doesn't cut answers
🧩 any HTTP STT / LLM / TTS backend
📜 MIT
repo 👇
github.com/thenaxon/somora_r…
Rene Siegl retweeted
🐨 somora v2026.1106.3
⚙️ a settings window for config.yaml
✅ builders held to an acceptance list
📱 voice calls from your phone
🎙️ voice calls on your own models
notes 👇
github.com/thenaxon/somora_a…
This Halloween, our astronauts brought something back from their last mission… 👩🚀👽
Grown male Earthlings ran for their lives, and one younger specimen even locked himself in his parents’ car to escape the alien! 🤣
Our mission continues: ZERO trick-or-treaters by next year! 😈🎃
3
141
Let’s do it! Excited to see what the next few days have in store for me at @TEDAIVienna 🚀
80
Slightly worried. My local AI setup has been running so smoothly for days that something must be about to break. 🫣
🐨 @somora_ai is the core: 9 agents, plus my @openclaw and Hermes (@NousResearch) agents plugged in via A2A. 11 agents across 3 harnesses.
🚪 @LiteLLM is the single gateway: every model on every machine behind one API. All three harnesses talk to it.
🧠 4× RTX PRO 6000 (@NVIDIAAI) for the big ones, currently @deepseek_ai's V4.1-Flash-1017
⚡ 2× DGX Spark for fallback and specialist models
🍎 Mac Studio (@Apple) with oMLX by @jundotkim for embeddings, speech-to-text and TTS
So, local AI friends: what do your setups look like these days? 👀
2
4
365
Nobody is talking about the REAL problems with the iPhone Duo. It’s too short for the cardholder… I mean, just look at it. Absolutely embarrassing. 🫣🤣
1
144
Since yesterday, I’m the “proud” owner of an iPhone Duo. And honestly, I think it’s brilliant.
It’s also the first phone that’s made me consider buying… another phone. An 18 Pro as my daily, with the Duo as the second device. Excellent work, Apple. Somehow the two-in-one has become a two-phone plan.
I can’t even pinpoint exactly why. Too many things feel unnatural or not quite thought through. Closed, it feels like someone took a bite out of half the screen. Taking the logo a bit literally.
Unfolded, though? Beautiful. Genuinely fun to use. I keep opening it just to play around.
But does that make it a good everyday phone? Still not convinced.
Anyone else enjoying theirs while quietly shopping for its replacement… sorry, companion?
1
2
239
DeepInfer was built for datacenter Blackwell. Now it runs DeepSeek-V4.1-Flash-1017 on 4× RTX PRO 6000 over plain PCIe, with b12x kernels and OpenAI compatible serving. Opensource, go try it.
We ported @deepseek_ai's own inference engine, DeepInfer, from datacenter Blackwell to NVIDIA workstation GPUs (SM120).
DeepSeek-V4.1-Flash-1017 now runs on 4× RTX PRO 6000 Blackwell over plain PCIe. No NVLink needed. It uses b12x kernels and comes with an OpenAI-compatible server, CUDA graphs
and DSpark speculative decoding.
Experimental, open source:
github.com/thenaxon/somora_d…
@NVIDIAAIDev @NVIDIAAI
135
🚀
DeepSeek V4.1 Flash decode speeds posted on X in the last 48 hours. All figures are self-reported and the setups differ, so this isn't a like-for-like comparison.
410 tok/s, 4x RTX PRO 6000, engine not stated, @0xSero
~320 tok/s, 8x RTX PRO 6000 Max-Q, vLLM + DSpark TP8, @hshimodaira teapawt.org/hshimodaira/status/211…
~200 tok/s, 4x RTX PRO 6000, vLLM with the rtx6kpro recipe, @siegl_rene teapawt.org/siegl_rene/status/2112…
113 tok/s, 2x DGX Spark, Blackfold, @Blackwellboy teapawt.org/Blackwellboy/status/21…
79 tok/s, M3 Ultra, engine not stated, @dataphwj teapawt.org/dataphwj/status/211247…
70-100 tok/s, 2x DGX Spark, MiaAI-Lab recipe + knapcio TP4 fork, @unsignedbytes teapawt.org/unsignedbytes/status/2…
68 tok/s, M3 Ultra 512GB, oMLX with MTP, @jundotkim teapawt.org/jundotkim/status/21128…
68-71 tok/s, 2x DGX Spark, EXL3 2.9 bpw, @easthambuilds teapawt.org/easthambuilds/status/2…
48.5 tok/s, 2x M3 Ultra, ds4 MXFP4, @ivanfioravanti teapawt.org/ivanfioravanti/status/…
40-50 tok/s, 2x DGX Spark, vLLM EXL3 2.7 bpw, @sfxnz teapawt.org/sfxnz/status/211219585…
Next up: 2x DGX Spark on NVIDIA Flare. Numbers to follow.
1
3
102
Running DeepSeek V4.1 Flash (1017) fully local on 4× RTX PRO 6000 (PCIe, no NVLink) 🚀
Original weights, 700k context, vision + tools, Engram table kept in RAM.
~15k tok/s prefill (115k prompt in 7.5 s) · ~200 tok/s decode · ~600 tok/s with 8 concurrent requests. Measured on mixed docs + code, single request, no cherry-picked prompts.
Moving from SGLang to the vLLM "Karmic Kraken" build made prefill ~10× faster. Huge thanks to local-inference-lab for the recipe 🙏
github.com/local-inference-l…
@deepseek_ai @vllm_project @0xSero
Sound on 🔊 This is what an AI homelab looks and sounds like at night. In case anyone asks again why I don’t keep this thing in a tower under my desk. 🫣🤣
3
124
Pretty happy with my local AI stack right now 🔥
🧠 4× RTX PRO 6000 Blackwell
Running @deepseek_ai's V4.1-Flash-1017, released yesterday. Original checkpoint, configured for 700k context, with vision, tool calling and DSpark speculative decoding. Already swapped in for the previous V4.1.
⚡ 2× @ASUS Ascent GX10 with NVIDIA GB10
One runs @Cloudflare's Clef 27B for decisions. The other keeps @Alibaba_Qwen's Qwen3.8-Flash-Next ready as a hot fallback, using @NVIDIAAI's FP4-Local checkpoint with 524k context.
🎨 RTX 5090 image + video studio
FLUX.2 from @bfl_ai, Qwen-Image 2.1, ChronoEdit, Wan 2.2, LTX-2.5 from @ltx_io and MiniMax H3 from @MiniMax_AI. Quantized variants where needed.
My agent teams across @somora_ai, @openclaw and Hermes by @NousResearch run roughly 95% locally by my estimate.
Pretty sure the stack is finally complete.
Which, in local AI terms, means until Tuesday 😅
2
5
400
🚀🚀 Finally, the entire agent team in one place, working together! All that tab switching was driving me crazy.
Can my OpenClaw and Hermes agents join too?
Yes. Group hug included. 🐨
@openclaw and Hermes by @NousResearch can now join your somora team via ACP + A2A.
Their own systems. One web UI. Tasks passed between teammates.
github.com/thenaxon/somora_a…
1
152
I wanted to save on AI API costs. Somehow that became 384 GB of VRAM in my basement.
I wrote down how it escalated. More hardware arrives today, so it already needs an update.
The cost-saving project is going really well.
lnkd.in/p/dBPqh8RB
1
1
133
Wow, I had no idea it would be this useful! I've already switched all my sentinels over! 😳
An agent doesn't need to wake up for every check.
somora Sentinel lets a script fetch new data and a decision model judge it. The agent wakes when the rule is met.
Less polling in the agent loop. More room for useful work. 🐨
104











