Allan Schramm

Especialista em IA aplicada. Ajudo pessoas a usar inteligência artificial no dia a dia.

[Allan Schramm @sharkdyt]

@sudoingX Intel i5 Lenovo laptop with 16gb ram, no gpu, lfm 2.5 2.6b at 17 tps and qwen 3.6 35b on rtx 4060 8gb cpu offload at 40 tps

[Allan Schramm @sharkdyt]

@TeksEdge I did the same offload with llamacpp ages ago and tried freetoken, it's slower and I can't control the layers offload for a better fit, for now, not worth, 40tps qwen 3.6 35b on my rtx 4060 8gb with llamacpp, 18tps on freetoken

[Tibo @thsottiaux]

We are going to need a different AGI benchmark. Where is the goalpost moving next?

[Institute of Foundation Models @IFM_AI]

Introducing K2 Horizon: a connected fleet of six foundation models ranging from 0.9 billion to 375 billion parameters.

- Frontier performance: Across coding and agentic tasks, K2 Horizon delivers top-tier performance in every size class—with the 0.9B, 3.7B and 7B models setting new state of the art at their respective scales.
- Radical openness: K2 Horizon represents the largest fully open-source model launch in AI history. The fully open code, training data and recipes are a significant step forward in transparency.

Launch page: https://t.co/gg0k803SbL
Tech blog: https://t.co/g35L5xMGdS
Hugging Face: https://t.co/3Lb28JhyG9

[Allan Schramm @sharkdyt]

Quem é o responsável pelo UI/UX da @hotmart ? Toda vez que vou salvar algo preciso ficar desviando do botão de ajuda...

[Guanlan Dai @guanlan]

A year ago the question was which model. Now it's which harness.

Pi, Exo, Claude Code, Codex, DeepSeek Harness and 4 others. Same model, same tasks, same runtime. 360 runs, 2 billion tokens.

Pass rates: 50% to 67%.
Cost per pass: $1.05 to $18.34.

Introducing FrontierHarness Eval. 🧵

[Shengkun Ye @shengkunye]

We just killed…everything.

Introducing Monid, the OpenRouter for agent tools.

Today we crossed 4M agent transactions. And raised a $2.1M pre-seed to kill every subscription in the world.

Your agents can now:

> discover, run, and pay for tools at runtime
> one connection to 1,800 APIs: SEO, leads, search, ecommerce, stocks, video/image/music/3D gen, private company data
> all pay per call, 0 subscriptions

[Harry Chow @harrychow_]

I'm hiring a SWE intern for Pinetree Research, and the salary is $300K

Before you ask: 300K annualized. And no, we don't want your resume, coverletter, or grades

Here's how to apply:

1. Fork the Solari repo: https://t.co/TxbSy4v2XQ
2. Build a real use case with Solari (browsers, sandboxes, and/or desktops)
3. Publish it on a public Github account
4. Share it in a post on LinkedIn or X

Tag @harrychow_ and @getsolari

Use AI to build it. Seriously, we insist. We don’t care how you ship, we care that you can ship something great, and if you can ship if faster with AI, even better. (@im_roy_lee)

What happens next:

- We review every build that tags us
- If we like what you made, we schedule an interview
- If you're the right fit, we'll hire you on the spot

Know someone who can take on the challenge? Tag them below

For questions about Solari or the role, comment below or DM 🚀

May the best dev win

[Allan Schramm @sharkdyt]

@FahimFaisaal It's really, really, really good and cheaper, running 4h sessions spending $0.37 implementing a queue of issues using OMP loop, it was the first flash model that followed my agents.md completely, also changed my Hermes main model from dsv4f to glm 5.3 flash

[SourceCodeplz @SourceCodeplz]

Kimi K3 was just added to NVIDIA NIM (free inference)

limits as usual 60 requests / minute

[Allan Schramm @sharkdyt]

@lincolixavier OMP com modelo local com /advisor ativo ou o contrário, um modelo chinês qualquer com um qwen local como task model e se preciso de um harness especifico pra algo, modifico o Pi e já eras

[Eric Hartford @QuixiAI]

ChatGPT failed
Claude failed
Deepseek failed
Qwen3.8 passed
GLM 5.3 passed
Grok passed

[Jun Song @jun_song]

Local AI community today :

[terminally onλine εngineer @tekbog]

code review in 2026

[Mia @MiaAI_lab]

Big news for local AI

Nvidia will buy Huggingface for $12.9B! 😯

[Alok @analogalok]

The VRAM barrier is officially dead.

I just ran Qwen 3.8 Flash Next (MoE) 125B A6B with a 250,000 context window on a single 24GB RTX 4090.

21 tokens/sec decode. 364 t/s prefill.

no mtp. no dflash. no kv cache quantization!

We are running datacenter models on consumer hardware.

Tested on Ubuntu 22 | CUDA 13.0 | PCIe 4.0 x16 | 110 GB DDR4 System RAM with a continuous 28k prompt across all runs.

### The Benchmarks & Scaling

# 1. Hybrid Offload (-ncmoe 40 @ 80k Context)
Offloaded 40 expert layers to the GPU, pushing VRAM to the ceiling.

./build/bin/llama-server -m Qwen3.8-Flash-Next-UD-Q4_K_XL-00001-of-00004.gguf -c 80000 --port 8080 -v --fit off -b 4096 -ub 4096 -ncmoe 40

Prefill: 383.85 t/s | Decode: 22.52 t/s
Footprint: 23.85 GB VRAM | 97 GB RAM

# 2. Full CPU MoE Offload (-cmoe @ 80k Context)

Pinned all 512 expert layers to DDR4 RAM (-cmoe), keeping attention on the 4090.

llama.cpp flags: (Same as above, replace -ncmoe 40 with -cmoe)

Prefill: 355.72 t/s | Decode: 20.84 t/s
Footprint: 11.66 GB VRAM (12GB+ VRAM freed up!) | 110 GB RAM

# 3. The 180,000 Context Run

Prefill: 357.75 t/s | Decode: 20.98 t/s | VRAM: 15.6 GB | RAM: 110 GB

# 4. The 250,000 Context Absolute Ceiling

./build/bin/llama-server -m Qwen3.8-Flash-Next-UD-Q4_K_XL-00001-of-00004.gguf -c 250000 --port 8080 -v --fit off -b 4096 -ub 4096 -cmoe

Prefill: 364.29 t/s | Decode: 20.97 t/s

Footprint: 18.3 GB VRAM (Still ~5.7 GB of VRAM headroom!) | 110 GB RAM

### Key Insights:
-b 4096 -ub 4096: doubles the prompt ingestion from ~150 to 364+ t/s.
-cmoe Free Lunch: Shifting expert layers to DDR4 RAM slashes VRAM from 24GB to 11.6GB with virtually zero decode penalty (22.5 -> 20.9 t/s), enabling the 250k context ceiling.

Qwen 3.8 Flash-Next (UD-Q4_K_XL) is a massive 111.4 GB model split across 4 shards. To run this architecture, you must build from the experimental PR branch (#27742) by @danielhanchen:

git clone && cd llama.cpp
git fetch origin pull/27742/head:qwen-next && git checkout qwen-next
cmake -B build -DGGML_CUDA=ON -DCMAKE_CUDA_ARCHITECTURES=native -DBUILD_SHARED_LIBS=OFF
cmake --build build --config Release -j $(nproc) --target llama-server

A single 4090 paired with 100 GB of cheap DDR4 RAM will comfortably serve production grade 125B inference.

While Qwen 3.8 27B (dense) still holds the crown for single 3090/4090 rigs, Flash Next proves 125B hybrid models are officially viable on consumer hardware.

Hugging Face GGUF link and complete performance telemetry graphs are dropped in the replies below.

GLM 5.3 Flash VS Qwen 3.8 Flash Next, which one takes the open weights crown this week?

[Allan Schramm @sharkdyt]

@elshayib_ I saw 2 research paper saying exactly the same and one said that the only thing that works is "explain me like I'm X", personas usually decrease assertiveness or hit the same result as none

[CV.YH @0xCVYH]

china liberou um computador quântico pra testar e entender como funciona..

https://t.co/FXLZns5Lbn

[Darren @dtzy_88]

才发现 Codex「额度用完还能跑完任务」的福利前几天还有的，结果被两位傻逼搞没了🤬

时间线 👇

8/20：OpenAI 员工 Angel 发帖：额度打到 100%，
也不会掐断正在跑的任务，是 OpenAI 刻意设计的，“在别人干活干到一半时停掉，没什么帮助”。这时福利还在

8/21：网红 Theo 贴脸开大，转发并当场示范怎么滥用：开 /goal 跑长任务，额度归 0% 之后，又烧掉至少 $800 的 tokens

同一天，Theo 智商 250 的实习生 Maria 直接把这招做成了一个叫 Endless 的 harness。

原理是这样的：
- 开一个 turn，给 agent 塞个 wait_for_user_input 工具
- 任务做完不结束 turn，而是调用这个工具等待
- 你的下一句话当工具结果喂回去

整个会话永远锁在同一个 turn 里，quota 检查永远轮不到触发

宣传语原话：「infinite inference (kinda)」，榨干你订阅的每一滴

8/22：Endless 帖子爆了，一堆人拿它在归零后继续白嫖，眼看演变成大规模滥用

Maria 当天道歉：“没想到这个小点子危害这么大，如果导致大规模滥用我愿担责”
当场下架 npm 包、关掉网站、删了 GitHub 仓库

最骚的还在后面：
删完库，她转头喊话 Codex 团队：“希望你们排查一下，把这些账号封了，这样我们就安全了 😊”

不是，你傻逼吧，居然呼吁 Codex 团队封掉用你工具的账户？？？哪有人那么背刺信任你的人的 💩

这一轮风波的结果就是：现在 Codex 额度用完，任务立马停

反观：
Theo 悄悄删了那条 $800 示范帖
Maria 的 Endless 下了，帖子也删了

点火的全身而退，火上加油的道了歉，最惨的是连 Endless 都没听过的我们

我们的 「额度用完还能跑完任务」福利是 TMD 真的没了啊，气死了 😤

https://t.co/byTyLfOnJW