Mapping Local Nodes – Mildlyinteresting
I've been working on mapping (with tags) and steering local models based on their activation path in specific context to questioning during a/b testing. There is…
Latest
Follow-up: DeepSeek V4 Flash on 2x RTX PRO 6000 finishes real coding tasks faster than Sonnet and Opus, at about Sonnet quality
This is a follow-up to post about which local models stay fast deep into long context and I learned a lot from people here. I kept…
Claude Code and China: The mechanism is activated when the user sets the ANTHROPIC_BASE_URL environment variable (used for local models)
The list of suspicious hostnames is not stored in plaintext within the code; instead, it is Base64-encoded and then encrypted using a simple XOR operation with…
Follow-up: GLM-5.2 NVFP4 on four DGX Sparks — the MTP mystery is solved, and it’s now ~24 tok/s at 128K context
Follow-up: GLM-5.2 NVFP4 on four DGX Sparks — the MTP mystery is solved, and it's now ~24 tok/s at 128K context This is a follow-up to…
Pay attention: a few chats waiting in tray reserve 1GB VRAM for themselves.
If an application uses a Web-based interface and "hardware acceleration", it constructs its frame in VRAM and sometimes keeps it reserved even if the app is…
Toolport: Use as many MCP servers as you want without the token tax
Disclaimer, this is a self promotion post but I truly believe it belongs here and is quite useful and relevant to this crowd specifically. I built…
Can we use SLMs to compress data?
Forgetting the villainization of overfitting, what happens when you just take a massive amount of training data and overfit the model on that. Is it possible…
[audio.cpp] The Sound of GGML — C++/GGML native ACE-Step, Stable Audio, HeartMuLa, RoFormer, HTDemucs released. 10-Minute Music in 60 Seconds!
https://preview.redd.it/yxa9dlzquxah1.png?width=2000&format=png&auto=webp&s=b07c74b8832b26b46531e2fddba19fd2437ce4c6 I just released a big music/audio expansion in audio.cpp. This batch adds music generation, SFX generation, and source separation to the released framework surface: Newly…
I benchmarked RAG techniques on a synthetic healthcare database. The biggest gains came from document shape, not model tweaks.
I built a small RAG benchmark over a synthetic clinic database because I wanted to test the usual advice instead of just repeating it. Benchmark The…
Hierarchos: Preliminary Findings From a 232M Recurrent Memory-Augmented Assistant Model [P]
Project Release / Research Draft] Hierarchos at 232M Parameters: Preliminary Findings From a Recurrent Memory-Augmented Assistant Model Technical Report: July 2nd, 2026 Project: Hierarchos / KortexHOS…