DEV Community: Andrale

Charming My Biblically Accurate Mamba

Andrale — Fri, 03 Jul 2026 18:38:49 +0000

So how to find the mamaba for you,

Standard mamba model; honestly pretty great but is it worth the hype? Absolutely have u seen how much transformers get? Narcs
Hybrid models; alr now we're talking honestly mamba-1.58 version is a wet dream if u don't care about precision as much, linear attention (keeps ur tokens and tokens per second [t/s] in check), u can also use zamba2 I feel it's legendary,
Ultra Low ram users like me; best bet is sub-1B if pure ram,and u want good speed,and not want ur device to go bye-bye and kamakaze
Power users who r rich 🤑; the world is ur to take it recommend liquid AI, it's multimodal as well, if u feel extra lengthy and girthy, try learning to graft or auto distill from a larger model like deepseek R1 so this babe actually can do autonomous tasks, the reason is precision, SSMs are notorious for batshit precision, tho mamba3 does solve most of it
Overall good family, very open source, and barely any incest

HDC model for dummies (18+)

Andrale — Fri, 03 Jul 2026 18:05:54 +0000

So well come to my guide for HDC(hyperdimension computing)for beginners and dummies, so ur here to learn HDC good, very maso, love it, 1.first u need python(1.0+)
2.next for libraries it's pytorch + torchdd or holo(If ur feeling extra kinky) which is a rust-python wrapper,

AI example code import torch import torchhd

1. Set the dimensionality

d = 1000 # The "1000D" you mentioned

2. Create a random hypervector (your first "memory")

hv = torchhd.random(1, d) # A tensor with 1,000 random values
print(f"Hypervector shape: {hv.shape}")

3. The fundamental operations

hv2 = torchhd.random(1, d)

BUNDLE (superposition): Think of this as creating a "set" or a "noisy" memory

bundled_hv = torchhd.bundle(hv, hv2)

BIND (association): This creates a new hypervector that encodes a relationship

bound_hv = torchhd.bind(hv, hv2)

4. Similarity (the "search" and "recognition" function)

similarity = torchhd.cosine_similarity(hv, hv2)
print(f"Similarity between two random vectors: {similarity.item():.4f}")

You'll see a number close to 0, as random vectors are nearly orthogonal.

(I recommend claude for free and gemini pro if ur rich 🤑to code for you if ur feeling extra romantic)

I recommend starting at 1k D, cos it's easier to control and won't get you scratching ur head but itching for more
I recommend thinking them of a sheet of graphene(2D) to help u visualise, if ur feeling smart try thinking them of diamond lattice(3D), if u feel even spicier image them as a universe with galaxies as moving nodes cos "distance" is now an abstract concept
Stay hydrated
Have fun 😊
Ask questions, cos it's better if ur responsive and ready to learn -teehee ✨😘

Just another set up guide for a 4GB ram potato

Andrale — Fri, 03 Jul 2026 16:38:12 +0000

So to keep it short, I'll lay out the details for quick readers, it's Bitnet 1.58 bonsai-8B+ bitnet.cpp(or llama.cpp of ur lazy) + tools like persistent memory and auto batching,(or just use ollama and use community plugins) ummm yeah that's it, if ur in for the juice here's more:

Llama.cpp is good honestly works well for Bitnet unless u feel maso u can use Bitnet very similar low learning curve, but try use 512 batching (works for me)if u have a dedicated GPU unlike please use that, it'll get ur bestie,
If ur feeling risky use early speculation like a small 0.5B model but Bitnet is fast enough already(also adds unnecessary ram overhead, or idk lora TTT is a good way? Too many things to do)
Why Bitnet, speed and just raw general IQ is dense AF(gives like 7B accuracy and 45t/s but don't take my word for it,but also so I don't feel bad expect 25t/s)
Should or can u find something better? Absolutely
Maybe ask me questions, I'll answer in a few mins prolly(i hallucinate too)
Any upgrades u can add? In place TTT makes u a mad max model, but lora TTT is going crazy especially if u use the prototype Qlora + inplace TTT, next is like tool calling use TTT or lora to teach it permanently (remember to save if ur using TTT) yeah there's more but honestly this should get u going pretty smoothly