"developers" by @pkgsrc
Twitter accounts of pkgsrc developers
What am I looking at?
Oh nothing, just smolBSD running Xorg over VirtIO GPU with virgl enabled.
WIP here: github.com/NetBSDfr/NetBSD-sโฆ
picoclaw 0.5.2 is out! and smolClaw is up-to-date with it :) smolbsd.org/smolclaw/ thanks for your work @afjcjsbx !
iMil ๐ช๐ธ๐ฆ retweeted
PicoClaw v0.5.2 just dropped! โก๏ธ
Boosted prompt cache hit rates with usage reporting, WebUI MCP dashboard controls, sync context compaction, and memory tuning for low-spec boards.
#OpenSource #AI #Assistant #Golang #RISCV
iMil ๐ช๐ธ๐ฆ retweeted
Seems that well known freshports.org/ has interesting addition called daemonports.org ... and its for Ports of all BSDs :)
iMil ๐ช๐ธ๐ฆ retweeted
Now I also confirmed that 2080ti unlock is successful with similar approach as the 30-series' method
it reach nearly same speed as 350w 3090(locked) with only 250w in mmapeak and mamf-finder
github.com/KohakuBlueleaf/Koโฆ
Confirmed that 3090 unlock is successful
with 360watt power wall, I got 75T -> 135T (measured in Pytorch)
Which is insane
I used github.com/Zagitta/gpu-unlocโฆ with commit a9b0c75
Oh no! I have created a frankenmonster of Cargo and autotools and other weird oddities like mtree. ๐ซ
Why? You'll soon see once I have time to iron out a first last issues...
// TODO: Insert "I'm not proud of it. I'm a bit." meme.
ALT Screenshot of the project's top-level directory (via ls), with highlights on Cargo.toml, Makefile.am, configure, and mtree.
Look! Just loads of modern device/fs drivers for old NT versions. Oh, they are vibe-coded of course. "What if they don't work perfectly?!?" But does it matter? You'll have retro-fun anyway. github.com/ages2001?tab=repoโฆ
ALT Screenshot of the quoted page showing the first top 3 repos.
iMil ๐ช๐ธ๐ฆ retweeted
Today we're launching cve.now ๐จ
Yesterday was a rough day for security. php2shell, LiteLLM, Railway... How are we supposed to stay on top of all of this?
cve.now aggregates 4 vulnerability sources into one free feed and uses AI to explain each one in plain English. No login.
cve.now
My now-traditional(?) BazelCon 2026 recap is here! jmmv.dev/2026/10/bazelcon-20โฆ . I'll warn you: it's _long_ and it talks about AI but... fear not, it's not AI-generated. Enjoy!
Give this project more visibility tweeps! Possibly the best personal assistant around, not overly complex, simple to install and run, works on anything, RT @afjcjsbx !
Do you want a personal AI assistant to integrate on WeChat, Feishu, Telegram? you can take a look at PicoClaw
A super lightweight open source agent with long-term memory, planned tasks, MCP tools, plugins and web interface
Installable with a single command, just enter the API key of the model and use it
github.com/afjcjsbx/picoclaw
iMil ๐ช๐ธ๐ฆ retweeted
Do you want a personal AI assistant to integrate on WeChat, Feishu, Telegram? you can take a look at PicoClaw
A super lightweight open source agent with long-term memory, planned tasks, MCP tools, plugins and web interface
Installable with a single command, just enter the API key of the model and use it
github.com/afjcjsbx/picoclaw
iMil ๐ช๐ธ๐ฆ retweeted
Si vous vous ennuyez au travail, sachez que vous pouvez dรฉsormais jouer ร Need for Speed Underground 2 depuis votre navigateur ๐ il y a mรชme du multijoueur !
Lien : nfsu2.xyz/
Hear me out: @support_huihui abliterated Qwen Flash Next 3.8 Swift iq2_xs on Strata running on a *8GB RX 7600*, 64GB DDR4 and a cheap NVMe drive at FREAKING 30-40 TOKENS/S
@coldniko is the goat.
Some neat & useful(!) swag I brought back from BazelCon ๐ I have something even cooler to show but that'll have to wait until I write an article about it. h/t @engflow
@_adriend_ toi qui est dรฉjร versรฉ dans le monde des LLMs, tu as essayรฉ Strata github.com/Niko1221/Strata ? Les rรฉsultats sont incroyables...
iMil ๐ช๐ธ๐ฆ retweeted
Strata 0.2.0 is now live in beta! It includes its own in-app harness and terminal-serving interface.
github.com/Niko1221/Strata
A lot of performance improvements and support for low-end devices ๐
iMil ๐ช๐ธ๐ฆ retweeted
PicoClaw v0.5.1 just dropped! โก๏ธ
โFeaturing OpenCode Go provider support, full Plugin CLI, smart artifact spilling for oversized outputs, and reduced memory usage on embedded boards.
โGrab the release:
github.com/afjcjsbx/picoclaw
โ#OpenSource #AI #Golang #RISCV
iMil ๐ช๐ธ๐ฆ retweeted
Real iOS 16.5 booting on Android hardware, Not a theme, Not an emulator.
HackiOS shows a Poco X5 Pro/ Note 12 Turbo booting into what looks like iOS 16.5 (Settings About as iPhone X, home screen, widgets).
They say custom UEFI + XNU kernel not a skin.
But no public source code, no independent verification, and no installable build.
They claim custom UEFI + XNU adaptation and say itโll be free and open.
iMil ๐ช๐ธ๐ฆ retweeted
La semaine derniรจre on a fait connaissance avec #FreeBSD. Alors demain, รงa va devenir un aquatique et piquant. Il est temps qu'on parle tu gros posson tout zaune ! #OpenBSD install / packages / ports / build... Et comme toujours, รงa se passe ici ร 8h: twitch.tv/lefinnois_
Writing articles in Markdown comes with friction. One thing that has always irritated me is inserting cross-links to other articles. I don't know elisp and would have never automated this... but AI does so... why not ask it? 5 minutes, and done. "SPC #" for my own blog helpers!
i don't get the moe obsession, every model release this week is moe
the last good dense model that fits on a decent gpu was qwen 3.8 27b in august, before that gemma 4 31b in april. that 20 to 30b dense class is where small gpus live, and it's gone quiet
and on small gpus dense is the one that pays off, give us dense models that run well on small gpus.
Maybe try QFN for real? It literally destroys 27B on every task. It runs on a 8GB GPU w/ 64GB DDR{3,4,5}
Qwen Flash Next UD-IQ4_XS is good, but note the level of detail UD-Q4_K_XL brings in this small X-Wing simulator, created in 10 minutes with Strata
* xs imil.net/stuff/xwing.html
* k_xl imil.net/stuff/xwing-q4_k_xlโฆ
A principle thatโs hard for me to not apply, for better or worse:
Faced with a problem, donโt just work to solve it. Instead, figure out what would have made it so easy that someone would have already solved it by now. Make it so. Then be the lucky someone.
I've been very quiet around here for unclear reasons, but with BazelCon 2026 behind our backs... I need to write a couple of articles so let's pick this up again!
@Rapid_API @nokia One month since your September 15 reply, still no response or provider payout despite several follow-ups. Could you please provide an update?
I've just PR'd AMD GPU support to @Vec_Szn's fork of Strata, unleashing 60 tok/s on Qwen 3.6 35B A3B with an RX 7600 / 8GB! github.com/VecSzn/PolyStrataโฆ
iMil ๐ช๐ธ๐ฆ retweeted
Hopla! Le replay du stream de dimanche, "Dรฉcouverte de #FreeBSD" est en ligne sur Youtube -> youtube.com/watch?v=yhq4ga5tโฆ
Current: [strata] done: 6719 tokens in 66 s (102.4 tok/s) (disconnect, cancel=True), expert cache 97.5% hit (+0.4% of the routed experts over PCIe)
The model is qwen3.8-flash-next-unsloth-ud-iq4_xs on 1xRTX5080 + 1xRTX3090 + 64GB DDR4
insane.
kudos to @coldniko
iMil ๐ช๐ธ๐ฆ retweeted
People still dont fully grasp Strata
You can now buy old cheap DDR3 machine with 128GB or 64GB RAM connect 8GB+ GPU and run Qwen3.8-Flash-Next at 70tps+
Remember guys: More cores/threads in CPU - faster it will go :)
github.com/Niko1221/Strata
smolClaw is now updated with @afjcjsbx's picoclaw fork release 0.5.0! smolbsd.org/smolclaw/
Great job Mauro!
iMil ๐ช๐ธ๐ฆ retweeted
PicoClaw v0.5.0 is here! โก๏ธ
โHuge update with Agent Plugins 1.0, MCP OAuth, Voice Mode, Image Generation, WebUI conversation forks, and a new low-power profile for embedded hardware!
โFull changelog & release:
github.com/afjcjsbx/picoclaw
โ#OpenSource #AI #Golang #RISCV
1. reduce source tree to strictly needed files
2. remove unneeded code paths
3. while at it, make it faster
Local inference BTW, Qwen 3.8 Flash Next iq4_xs on Strata github.com/Niko1221/Strata, all thanks to @coldniko
iMil ๐ช๐ธ๐ฆ retweeted
- GPU Kernels and Type-Specific Computations: Completely rewritten in SYCL (cannot be merged).
- Multi-GPU Workload Distribution: I also plan to implement distributed execution across multiple computers over a network in the future (cannot be merged).
iMil ๐ช๐ธ๐ฆ retweeted
๐จ๐ซ๐ท La taxe GAFAM passe de 3% ร 15%.
L'Assemblรฉe prรฉsente รงa comme une "taxation des gรฉants amรฉricains" qui ne touchera pas les franรงais.
Sauf que...
C'est faux. ๐
Google, Meta, Amazon n'ont aucune raison d'absorber une hausse de taxe de 12% sur leurs marges.
โ Ils vont la rรฉpercuter sur les annonceurs qui achรจtent de la publicitรฉ Google ou Meta (PME franรงaises, artisans, e-commerรงants).
โ Sur les vendeurs Amazon qui paient leurs commissions.
โ Sur les abonnรฉs aux services premium.
Un impรดt sur une entreprise dont le pricing est libre de s'ajuster est toujours, ร terme, un impรดt sur ses clients.
Entre 3,5 et 4 milliards d'euros de recettes : les dรฉputรฉs votent en commission pour porter la taxe Gafam de 3% ร 15%
l.bfmtv.com/li5g
iMil ๐ช๐ธ๐ฆ retweeted
Was always a big fan of @gospaceport โฆ now Iโm seeing my project in his video ๐ฅฒ
youtu.be/x7s6Mf4xF9A?is=dIDBโฆ
Incredible.
On an RX 7600 / 8GB + 32GB DDR4 ROCm 10.1.0, qwen3.6 35b a3b q4_xs with llama.cpp is *slower* (~15 t/s) than qwen 3.8 QFN iq1_m with Strata (20+ t/s) cc @coldniko
iMil ๐ช๐ธ๐ฆ retweeted
People buying Mac Minis to run OpenClaw while the #PicoClaw gateway takes just ~38 MB of RAM (with only ~18 MB of private memory!). โก๏ธ
And this isn't even a clean install, it's driving my daily setup with 6 MCP servers and 1 plugin.
Peak efficiency! ๐ฅ
#OpenClaw #AI #Assistant
iMil ๐ช๐ธ๐ฆ retweeted
Since PicoClaw by @SipeedLab seems unmaintained, Iโve forked it to keep it alive. Iโve already shipped several new features!
โWeโre looking for users (feedback/bug reports) and contributors to join in.
โgithub.com/afjcjsbx/picoclawโฆ
โ#OpenSource #AIAssistant #AI
Rejoice picoclaw users: despite @SipeedLab silence, @afjcjsbx is now maintaining an active fork of the cleanest claw around, 0.4.0 available at github.com/afjcjsbx/picoclaw
Hey @afjcjsbx ! Thanks for taking over picoclaw <3 I've seen your messages, just hadn't the time to re-open the PR on your fork but I'll certainly will, and possibly another one about conversations mixing with previous ones.
Create an animated SVG of a colorful parrot juggling with bottles of gin
imil.net/stuff/pinpin_ud_iq4โฆ
~55 tok/s on: RTX5080 + RTX3090, 64GB DDR4
Strata magic cc @coldniko
iMil ๐ช๐ธ๐ฆ retweeted
I know Niko's been busy, but I'm kind of impatient
This is one day of me with Claude Pro (cheapest plan) access๐
๐ฅGLM 5.3 Flash
โ๏ธ93.1GB model file (the smallest, for now)
๐About 20t/s decode
โ๏ธFully custom version of Strata
๐ฅNo MTP (yet)
๐Xeon E5 2690 V4
๐32GB ddr4 quad channel
๐NVME 3.0
โญ2x Tesla V100 PCIe 32GB
I have redated part of the dashboard numbers because it is all new and I don't want to spoil anything until that is ready. I left the experts location open though, in case anyone thinks that is not the model I claim :)
I enjoyed working on it, so I plan to carry on with optimisations and eventually release it all, stay tuned!
More to comeโณ
iMil ๐ช๐ธ๐ฆ retweeted
Hopla ! Le replay du stream "plus avant avec le LLM local" est en ligne sur Youtube (attention, l'intro peut contenir des rรขleries de vieux con rรฉac) : youtube.com/watch?v=XaK5d-Caโฆ
iMil ๐ช๐ธ๐ฆ retweeted
You can also donate via: buymeacoffee.com/strataenginโฆ
Funds will take two weeks though :(
iMil ๐ช๐ธ๐ฆ retweeted
Please share this to support Strata development ๐
Looking for sponsors or rental of at least:
- PC with AMD GPU + 64GB RAM
- PC with Intel GPU + 64GB RAM
- PC with Dual GPU setup + 32GB RAM
Location: Slovenia (Europe)
I would like to work faster as weโre only 30% in. Please message me
Thank you
Niko
TLD Zone File Statistics (Glue records, Name servers, DNSSEC, and Domain counts) for October 2026 are online statdns.com
iMil ๐ช๐ธ๐ฆ retweeted
Huihui-Qwen3.8-Flash-NEXT + Strata
ใฐใใธใงใ
huggingface.co/alesha-pro/Qwโฆ
omw @coldniko 85 tok/s w/ 0.1.39 vs 75 tok/s w/ 0.1.38 BEAAAST (RTX 5080 + RTX 3090 + 64 DDR4)
iMil ๐ช๐ธ๐ฆ retweeted
Booting up Version 7 UNIX on Misspiggy. Weโre open 10-4 today!
icm.museum
#vintagecomputer #retrocomputer #unix #pdp11
iMil ๐ช๐ธ๐ฆ retweeted
- Output t/s: 75.8 -> 80.0 (+8% aggregate)
- Prompt t/s: +26%
- Parallel requests (4 max) ๐
- Old CPU support: Sandy / Ivy Bridge, Xeon E5 v1/v2, Bulldozer, Nehalem / Westmere
- Old GPU support: P40, P100, GTX 10, V100, Titan V, Instinct MI50 / MI60, Radeon VII
- Intel Arc support ๐
iMil ๐ช๐ธ๐ฆ retweeted
Strata is great! Iโm particularly interested in improving multi-GPU performance.
I found that the helper cache stays static while the main GPU adapts, allowing duplicate experts and more CPU fallback. A small optimization delivered +62.5% mixed-task and +129.2% code decode speed on dual RTX 4090s.
Hope this can make it upstream! @coldniko
iMil ๐ช๐ธ๐ฆ retweeted
- Prompt processing on multi gpu +20%
- AMD performance increased by 12%
- Unsloth UD-Q4_K_XL experimental fix
- OrcaRouter IQ3_XXS added
- RTX 20 series fixes
- Many more improvements
github.com/Niko1221/Strata/rโฆ
Qwen3.8-Flash-Next first 32K tokens are 4%-10% faster, LOW RAM mode added (for low hardware).
New update here
github.com/Niko1221/Strata/rโฆ
iMil ๐ช๐ธ๐ฆ retweeted
Destroy any website with a stickman.
Have fun: destroy.spritefusion.com
@Iberia buenas! veo que se puede llevar una maleta de 10kg en cabina con billete economy, es verdad? Es un vuelo Valencia -> Ibiza