ModelForge

5 ms·

1.
Pacing != Pacing Development(sebastianraschka.com)
3 points by ModelForge 3d ago|0 comments
2.
GPT-6 Astra, looped transformers, and hidden reasoning(magazine.sebastianraschka.com)
519 points by ModelForge 8d ago|162 comments
3.
Claude Watermarks Text: Token sampling, watermark detection, and removal(magazine.sebastianraschka.com)
3 points by ModelForge 24d ago|0 comments
4.
How Claude's Text Watermarking Works [video](youtube.com)
2 points by ModelForge 29d ago|0 comments
5.
How Claude's Text Watermarking Works(sebastianraschka.com)
5 points by ModelForge 1mo ago|0 comments
6.
by ModelForge 2mo ago
Good question, it's 100%. I.e., the developers usually share a reference implementation with e.g., Hugging Face transformers to load their weights, and from there on you can read the code and, if you have time, reimplement and check ev
7.
by ModelForge 2mo ago
And adding to that, there is also the recurrent state in the Kimi Delta Attention. I wouldn't call it position information but more sth like "position sensitivity"
8.
Kimi K3 Architecture Overview and Notes(sebastianraschka.com)
507 points by ModelForge 2mo ago|111 comments
9.
Inkling: A New Open-Weight 975B Moe with a Few Surprises(sebastianraschka.com)
3 points by ModelForge 2mo ago|0 comments
10.
Claude Code's Real Secret Sauce Isn't the Model(sebastianraschka.com)
6 points by ModelForge 6mo ago|0 comments
11.
The State of LLMs 2025: Progress, Problems, and Predictions(magazine.sebastianraschka.com)
3 points by ModelForge 9mo ago|0 comments
12.
A Researcher's Field Guide to Non-Standard LLM Architectures(magazine.sebastianraschka.com)
2 points by ModelForge 11mo ago|0 comments
13.
Explanation of Gated DeltaNet (Qwen3-Next and Kimi Linear)(github.com)
3 points by ModelForge 11mo ago|0 comments
14.
The Core Components of Modern LLMs and the Models Beyond Transformers [video](youtube.com)
3 points by ModelForge 11mo ago|0 comments
15.
Popular Attention Alternatives: GQA, MLA, SWA(sebastianraschka.com)
4 points by ModelForge 11mo ago|0 comments
16.
Multi-Head Latent Attention(sebastianraschka.com)
4 points by ModelForge 11mo ago|0 comments
17.
Thinking Machines Lab Co-Founder Departs for Meta(wsj.com)
7 points by ModelForge 11mo ago|0 comments
18.
OpenAI's internal Slack messages could cost it billions in copyright suit(sherwood.news)
8 points by ModelForge 11mo ago|1 comments
19.
LLM Evaluation from Scratch: Multiple Choice, Verifiers, Leaderboards, LLM Judge(magazine.sebastianraschka.com)
4 points by ModelForge 1y ago|0 comments
20.
by ModelForge 1y ago
No the compiled version is actually faster. From that table, the A100 tok/sec (larger is faster) numbers are: - Eager: 28 - Compiled: 128 And - KV cache eager: 26 - KV cache compiled: 99 The reason that the KV cache is slower is likely
21.
by ModelForge 1y ago
Could be an artifact of the small size not fully taking advantage of the GPU. For example, for the slightly larger Qwen3 0.6B model the A100 is faster (you can see it when scrolling to the bottom here: https://github.com/ras
22.
by ModelForge 1y ago
I'd say the common ones (besides educational) are - private, on-device models (possibly with lower latency than models via web API); also edge devices - algorithm research (faster and cheaper to prototype new ideas) - cheap tasks, like
23.
Gemma 3 270M re-implemented in pure PyTorch for local tinkering(github.com)
417 points by ModelForge 1y ago|57 comments
24.
by ModelForge 1y ago
I think GPT-4.5 was potentially the original GPT-5 model that was larger and pre-trained on more data. Too bad it was too expensive to deploy at scale so that we never saw the RL-ed version
25.
by ModelForge 1y ago
The ollama one uses even less (around 13 GB), which is nice. Apparently the gpt-oss team also shared the mxfp4 optimizations for metal
26.
by ModelForge 1y ago
Good point. LLMs lower the barrier to entry if someone has enough resources because those architectures are more robust to tweaks given one throws enough compute and data at them. You can even violate scaling laws and still get a good model
27.
by ModelForge 1y ago
I’ve been using the ollama version (uses about 13 Gb RAM on macOS) and haven’t had that issue yet. I wonder if that’s maybe an issue of the llama.cpp port?
28.
GPT-OSS vs. Qwen3 and a detailed look how things evolved since GPT-2(magazine.sebastianraschka.com)
490 points by ModelForge 1y ago|97 comments
29.
LLM Research Papers: The 2024 List(magazine.sebastianraschka.com)
5 points by ModelForge 2y ago|0 comments
30.
Scaling Test-Time Compute with Open LLM Models(huggingface.co)
3 points by ModelForge 2y ago|0 comments

More ›