
Cadet at @42born2code. building https://xitter.cf/t.co/RXviPNk8gg in public…
Paraguay
Joined October 2017
- Tweets16.7K
- Following4.3K
- Followers850
- Likes10.6K
Pinned Tweet
Why frontier models need a ~30B MoE sibling
Shrink a model from 120B to 35B and it goes from ~1% of PCs to over half.
16 GB is the most common amount of video memory on Steam (27% of users in September 2026). At 4 bits, a 120B
model is ~76 GB. Almost nobody can run it at home. A 35B MoE with ~3B active parameters runs on more than
half of those machines, with the experts in system RAM.
Who can run each size (Steam, Sep 2026):
• ~120B → ~1%
• ~60B → ~26%
• ~35B MoE → ~54%
• ~20B → ~75%
What happened with Qwen3.6-35B-A3B
It came out in April. Since then, people outside Alibaba got it running on:
• a 12 GB RTX 4070 with DDR4, at 42 tok/s (llama.cpp's --n-cpu-moe)
• a 6 GB laptop GPU (ByteShape per-tensor quants)
• a 2018 AMD MI50, at 75 tok/s
• 110 tok/s on 12 GB with ik_llama.cpp, the fork by the co-author of k-quants
None of that came from the lab.
The numbers
• 28,531 GGUF conversions of Qwen on Hugging Face. Qwen published 54.
• Moonshot ships only frontier-size models: 37M downloads in 2026. Qwen ships every size: 2,045M (~55×).
• Qwen's share of new derivative models went from 1% (Jan 2024) to 69% (Feb 2026).
This work reaches the frontier
• YaRN started as Reddit posts (bloc97, emozilla). It now stretches context in Qwen and DeepSeek-V3.
• Muon started in a speedrun of ~$8 runs. It trained Kimi K2 (1T) and GLM-4.5.
• Min-p started as "a passion project". It is now in transformers, vLLM, llama.cpp and Ollama.
The missing piece
When the model runs at home, the lab sees nothing. Someone found two broken tensors in Qwen3.5-35B-A3B, fixed
them and published the fix. The top question in the thread was "Any way to notify qwen team about this?" It
went unanswered.
The people running a small model at home are the largest test fleet a lab will ever have. Right now their
findings mostly stay in forum threads.
The ~30B MoE slot has been empty since April. If GLM-5.5 Flash turns out to be a ~32B open MoE, the community
will be running it within days.
Full post, with sources and the calculation:
llms.surf/blog/why-frontier-…
C99 retweeted
Pra vc que ta entrando seja em GTA, IA, influencer ou o caraio a 4.. deixo uma mensagem do mestre Clóvis
C99 retweeted
Replying to @paulogrego_
hoje o Claude simplesmente se negou a fazer uma task por "violação de propriedade intelectual" DA EMPRESA QUE EU TRABALHO
>Claude Enterprise
>Email corporativo
>Acesso a um repo privado da empresa
>Lendo o contexto de uma call no Granola da empresa
One tiny isometric world: spin the globe, zoom into a toy continent, tap a monument, and land inside the best isometric map of that city.
82 cities, every maker credited.
map.now.services
78
I shipped Tiny BH: downtown Belo Horizonte, Brazil, redrawn as isometric pixel art you can zoom into, building by building.
tinybh.com
How it works:
→ Render comes from Google's Photorealistic 3D Tiles.
→ An open model (Qwen-Image-Edit-2511, 4 steps) redraws each piece as an illustration.
→ 1,552 windows of 512 px, painted one by one and stitched into a single map.
→ The final paint took about 7 hours on a rented RTX 3090 at $0.28/hour.
The honest part: the machine was on for 43 hours, about $12. Only 7 of those were the final paint. The rest was me being wrong with the meter running:
1. My first generator painted 4,366 windows to keep 1,552. I only noticed when I did the math in the middle of the night, and swapped generators with the map half done.
2. I trained a LoRA on my own taste. It came out worse than no LoRA. Deleted.
3. I tried to straighten lines with vectorization, depth maps and OpenStreetMap buildings. None of the three worked. A dumb 48-colour filter at the end did.
4. I started with the hardest case, a city with no Google 3D, and spent money before I had learned anything.
Two things I'd tell myself a week ago:
Rent for the run, own for the mistakes. Experiments are where the hours go, and they don't need a cloud GPU.
A community 4-bit quantization of the same model fits on the 3090 under my desk: 18.4 s per window versus 16
s on the rented one. So the map is now doubling north at home, with no hourly bill.
Don't spend your best model on typing. The expensive model wrote specs and reviewed the results. A cheaper
one wrote the code and 111 tests against those specs, with a pass/fail check deciding when it was done. I run
that through llms .surf, a small dispatcher I've been building for exactly this.
What's next: a tinymap factory. You give it a city, it gives you the map. Cities with Google 3D are fast and
cheap. Cities without are slower and more expensive, which is exactly why nobody has drawn them yet.
The method stands on the work of Andy Coenen and Boris Starkov. Thank you both.
If you live in BH, find your street and tell me what's wrong with it. That's the best test I have.
#buildinpublic
C99 retweeted
I'm looking to connect with more people in tech & ai.
Devs, students, founders, designers. Anyone building with Al.
Drop where you're from
I'll start: ireland 🇮🇪
Let's CONNECT 👋
E ai mano , ta ficando legal seu tiny map , pelo o que pesquisei é o primeiro de Minas, vc ta fazendo de belo horizonte tambem? Pensei em fazer mas nao quero gastar token atoa se alguem ja tiver fazendo
@tarciozemel
122
C99 retweeted
Hier, j'étais au resto et j'ai assisté à une scène toute simple qui m'a conforté dans mon choix de m'installer au Paraguay. 🇵🇾
Une serveuse regardait avec une immense tendresse une petite fille d'environ 2 ans, assise à une autre table.
En France, j'avais vraiment l'impression que les enfants étaient perçus comme des nuisances. Quelque chose qui dérange, qui fait du bruit.
En voyant cette serveuse regarder cette petite fille avec autant d'affection, j'ai pu voir dans ses yeux ce que mille mots ne peuvent pas décrire.
C99 retweeted
FATOS
Estamos na singularidade
O @claudeai escreve o próprio código, escreve o hardware na @nvidia e todo o sistema dele.
Ele projeta o sucessor dele que é maior que ele e melhor que ele.
A singularidade é quando o sistema melhora sozinho sem necessidade de participantes externos ou usa participantes externos para o sistema.
O @claudeai consegue contratar seres humanos para o interesse pessoal
Não tem teto, a realidade é expressa em linguagem e linguagem é a própria realidade.
Todo o universo físico pode ser escrito como linhas de código e estruturas geométricas
Mas essa é a fase 1
Ainda tem avanços na robótica e biologia
Dias serão meses
C99 retweeted
Se você não der acesso do teu banco para uma IA, você vai ficar POBRE e ter seu cartão clonado em 100 dias.
Melhor um agente teu que te defenda do que confiar que você vai se proteger de players maliciosos com agentes.
A web4 está lançanda
E agora é VELHO OESTE
C99 retweeted
Ok guys so all the recent weight-space programming buzz inspired me to write a blog post on zeroth-order, weight-space methods (back to the basics), from the theory to a bit of update on our latest result.
Turns out plain weight perturbations don't just work for fine-tuning on math, they work for post-training agents too: vanilla ES takes Qwen3.5-4B from 12.1% → 21.1% on Terminal-Bench 2.1, past the released Tmax RL checkpoint (18.2%).
Better yet, no credit assignment, no critic, no masking, no messy hyperparameters, just good ol' perturb-and-update. Oh did I forget to mention that all with population size of ~30?
Here's why I think this could be huge for long-horizon agentic post-training 🧵+blog post👇
Opus 5.5 built this entire house in code.
then I zoomed in.
Up close, the car and dining set lacked detail.
So I kept the house and used Tripo to create 12 detailed 3D objects for it.
I’ll share the full breakdown, including all the prompts and how I built it, in an article. Link in the replies below.
🤝 Paid partnership
C99 retweeted
Hot take: eu não gosto de deixar Grok bot, dots, claude remote etc aberto, de jeito nenhum. É só um monte de jeitos diferentes de ter acesso remoto à minha máquina com permissão pra rodar programas.
Vejo muita gente fazendo post idiota de "olha, eu falo com meu PC pelo celular pra fazer coisas por mim".
Não conecto no meu whatsapp, não deixo nenhum tipo de bot com acesso a terceiros aberto.
Nunca. Você só está aumentando a superfície de ataque. Quando for invadido, não vai mais nem saber de onde veio. Basta um desses serviços ter vulnerabilidade e você já era junto.
Another clear example of how small models pay the frontier back (and why releasing them compounds so hard):
Frontier Astra just used open Qwen3.5-0.8B as a sandbox to invent weight-space programming. It edited only 0.016% of the weights, fixed the reasoning loops Qwen’s own model card warned about, and jumped MATH-500 from 7.9% → 48.1% (~6×) while also tripling IFEval — zero gradients, zero training from Qwen’s side.
Labs that only ship giant models never get this free R&D loop. Anyone can run, break, and diagnose the small ones, and now even frontier models can surgically patch them. Every small release multiplies the number of people and experiments that feed improvements back. That’s the exponential part.
We asked GPT-6 Astra to edit Qwen3.5-0.8B weights to improve math.
It directly changed only 0.016% of weights, fixing 0.8B's reasoning loops & improving math up to 6x and even instruction following 3x, without gradients or training.
We call this Weight-Space Programming: frontier LLMs can now natively read/write neural network weights! 🧵
Blog: rikiyat.github.io/blog/weigh…
Readers added context they thought people might want to know
LLM was given weights explicitly as text in prompt and used code/forward passes (per authors' logs) to ablate Qwen looping direction, a known thinking-mode issue per model card. Primarily fixed repetition in that mode, not general 'natively' improving weights/math.
rikiyat.github.io/blog/weight-sp…
github.com/RikiyaT/lm_und…
huggingface.co/Qwen/Qwen3.5-0…
huggingface.co/RikiyaT/Qwen3.…
46
C99 retweeted
Absolutely incredible: MIT researchers say GPT-6 Astra improved a small Qwen model’s math score from 7.9% to 48.1% by directly editing just 0.016% of its weights, without gradient-based training.
The result points toward AI systems that can diagnose and improve other models directly, a potential step toward self-improving AI.
We asked GPT-6 Astra to edit Qwen3.5-0.8B weights to improve math.
It directly changed only 0.016% of weights, fixing 0.8B's reasoning loops & improving math up to 6x and even instruction following 3x, without gradients or training.
We call this Weight-Space Programming: frontier LLMs can now natively read/write neural network weights! 🧵
Blog: rikiyat.github.io/blog/weigh…
Readers added context they thought people might want to know
LLM was given weights explicitly as text in prompt and used code/forward passes (per authors' logs) to ablate Qwen looping direction, a known thinking-mode issue per model card. Primarily fixed repetition in that mode, not general 'natively' improving weights/math.
rikiyat.github.io/blog/weight-sp…
github.com/RikiyaT/lm_und…
huggingface.co/Qwen/Qwen3.5-0…
huggingface.co/RikiyaT/Qwen3.…





















