UX advocate, builder & agentic code writer. me on Nostr- npub12xyl6w6aacmqa3gmmzwrr9m3u0ldx3dwqhczuascswvew9am9q4sfg99cx ( Say no to crypto tokens or scams )
Pushin' pixels & fixin' UX
Joined February 2007
- Tweets31.7K
- Following3.7K
- Followers1.9K
- Likes95.1K
For anyone following along ... I have @deepseek_ai running locally & being driven as a 'sub-agent' via @claudeai ... the results match my real-coding experience...
Small jobs it's great ... bigger tasks ... DS needs a bit more management.
xitter.cf/vveerrgg/status/211548…
Do you want a hybrid @claudeai / local @deepseek_ai platform?
The idea is you ask Claude to "spin up sub-agents" & it uses your local instance & returns the results to your session.
... building on the work I did yesterday, I'll share it when I have it.
considering this is all open weights ... this lip-sync / music video tool is only going to get better from here ... @ComfyUI & @ltx_io #OpenSourceHiggsfield
Turns out ... this idea is VERY NEW in the world we're in.
I feel like this is something @TheAhmadOsman would be into.
NOW WERE TALKING !!! ... @Behringer ... release this beast!
Enable the software to accept API connections to platforms like @Suno in some sort of OpenGPT "like" format ... and I will build a generative music platform to match
Do you want a hybrid @claudeai / local @deepseek_ai platform?
The idea is you ask Claude to "spin up sub-agents" & it uses your local instance & returns the results to your session.
... building on the work I did yesterday, I'll share it when I have it.
Another day another lipsync bug to fix. … but as I’m learning. My coding methodology works GREAT at catching image errors as well.
Over render on the 1st pass & then clean & distill … rinse & repeat
#OpenSourceHiggsfield @ComfyUI & @ltx_io
It’s wild how fast anyone with a curious mind can go from ZERO knowledge… to useful insights by asking LLM questions.
Yesterday’s why & how questions have become today’s. Local LLM hosting at home.
Now thats a lip sync !!! ... and DEFINITELY should be part of the #OpenSourceHiggsfield contest with @ComfyUI
Another day ... another @github worm ... keep an eye on your code
Claude just saved my ass from a North Korean supply chain attack
Like actually
Thank you @AnthropicAI
keep an eye on your @github PRs ... looks like there's some bad code working around ... seeing some posts about it pushing PRs & forging user tokens.
@deepseek_ai V4.1 (748B) on stock llama.cpp. No Strata, no fork, no third-party engine. Repacked @ISTA_DASLab’s GSQ-RCO quant into upstream format: same weights, byte for byte.
Runs on 1 RTX PRO 6000 + RAM: ~40 tok/s.
🤗 huggingface.co/HumanjavaEnte…
🛠️ github.com/HumanjavaEnterpri…
Update: I ran a head-to-head vs @UnslothAI’s UD-IQ1_M (same size class, same box, same settings).
GSM8K: 97.2% vs 96.4% (tie)
MMLU: 84.4% vs 84.6% (tie)
Decode: 31.9 vs 24.3 tok/s (+31%)
Perplexity: unsloth wins (3.33 vs 3.51)
Raw per-question data is public 👇
github.com/HumanjavaEnterpri…
Tested the 2bit DeepSeek V4.1 repack against the full-precision model (via DeepSeek’s API)
Math: untouched (all three tie on GSM8K)
Knowledge: both 2-bit quants drop ~2.5 pts on MMLU
Compression blurs facts more than reasoning. Raw data 👇
github.com/HumanjavaEnterpri…
Couldn’t call it a night without a little extra tuning... turns out one flag (-ub 2048) made my other local DeepSeek read prompts 3.2x faster (51 → 166 tok/s).
And “threads = physical cores” isn’t a law: the IQ2 quant wanted 16, the Q4 wanted 30. Bench your own box.
Notes 👇
github.com/HumanjavaEnterpri…
what started as a Lip-Sync toolchain ... has become a bit of a music video Swiss-Army Knife.
" Middle shot: the mic (red) and her face (green), each found from one text phrase and followed through every frame. No hand-made masks."
-learning new tricks @ComfyUI @ltx_io #OpenSourceHiggsfield
Flowers to @Chaos2Cured 🌸 If Wednesday really is *the day* the internet breaks; at least I learned to roll my own local build of DeepSeek V4.1 on stock llama.cpp.
I’ve got a 748B model running locally. Even if this post might age like milk...
huggingface.co/HumanjavaEnte…