AI · alternatives · 2026

24 alternatives to Duplicate 3 layers in a 24B LLM, logical deduction .22→.76. No training

I replicated David Ng's RYS method (https://dnhkng.github.io/posts/rys/) on consumer AMD GPUs (RX 7900 XT + RX 6950 XT) and found something I didn't expect. Transformers appear to have discrete "reasoning circuits" — contiguous blocks of 3-4 layers that act as indivisible cognitive units. Duplicate the right block and the model runs its reasoning pipeline twice. No weights change. No training. The model just thinks longer. The results on standard benchmarks (lm-evaluation-harness, n=50): Devstral-24B, layers 12-14 duplicated once: - BBH Logical Deduction: 0.22 → 0.76…

This project demonstrates a technique for improving large language model reasoning by duplicating specific layer blocks without retraining. By duplicating contiguous groups of 3-4 layers in models like Devstral-24B… Below are 24 products that do a similar job, ranked by how close each is in meaning and then by launch-day votes.

1
How I topped the HuggingFace open LLM leaderboard on two gaming GPUs▲495
I found that duplicating a specific block of 7 middle layers in Qwen2-72B, without modifying any weights, improved performance across all Open LLM Leaderboard benchmarks and took #1. As of 2026, the top 4 models on that leaderboard are still descendants. The weird finding: single-layer duplication does nothing. Too few layers, nothing. Too many, it gets worse. Only circuit-sized blocks of ~7 layers work. This suggests pretraining carves out discrete functional circuits in the layer stack that only work when preserved whole. The whole thing was developed on 2x RTX 4090s in my basement. I'm…
Mar 2026 · dnhkng.github.io · its alternatives →
2
I built a tiny LLM to demystify how language models work▲915
Built a ~9M param LLM from scratch to understand how they actually work. Vanilla transformer, 60K synthetic conversations, ~130 lines of PyTorch. Trains in 5 min on a free Colab T4. The fish thinks the meaning of life is food. Fork it and swap the personality for your own character.
Apr 2026 · github.com · its alternatives →
3
Speeding up LLM inference 2x times (possibly)▲419
Here's a project I've been working on for the last few months. It's a new (I think) algorithm, that allows to adjust smoothly - and in real time - how many calculations you'd like to do during inference of an LLM model. It seems that it's possible to do just 20-25% of weight multiplications instead of all of them, and still get good inference results. I implemented it to run on M1/M2/M3 GPU. The mmul approximation itself can be pushed to run 2x fast before the quality of output collapses. The inference speed is just a bit faster than Llama.cpp's, because the rest of implementation…
2024 · asciinema.org · its alternatives →
4AB
AutoThink – Boosts local LLM performance with adaptive reasoning▲397
I built AutoThink, a technique that makes local LLMs reason more efficiently by adaptively allocating computational resources based on query complexity. The core idea: instead of giving every query the same "thinking time," classify queries as HIGH or LOW complexity and allocate thinking tokens accordingly. Complex reasoning gets 70-90% of tokens, simple queries get 20-40%. I also implemented steering vectors derived from Pivotal Token Search (originally from Microsoft's Phi-4 paper) that guide the model's reasoning patterns during generation. These vectors encourage behaviors like numerical…
2025 · its alternatives →
5
Netmind Power▲272
Hi HN, We just opened public access to Qwen3‑Coder‑480B‑A35B‑Instruct via a REST/SDK API. Why it might be interesting: - Fully open-source (Apache‑2.0): https://github.com/QwenLM/Qwen3-Coder - 480B params (35B active MoE), long context up to 1M tokens with YaRN - Matches Claude Sonnet/GPT‑4o on SWE-Bench Verified (81.6%) - Free tier: 100K tokens/month
2023 · netmind.ai · its alternatives →
6
ZeroGPU▲298
The compute efficient layer for AI inference
Jun 2026 · zerogpu.ai · its alternatives →
7
Fine-tune an 8B model on a 4 GB laptop GPU▲139
Fine-tune LLMs from one YAML. Layer streaming trains an 8B model on a 4 GB laptop GPU. - MakazhanAlpamys/Soup
Aug 2026 · github.com · its alternatives →
8
Minimal LLM Post-Training Experiments on an 8GB GPU (SFT, DPO, GRPO)▲21
Minimal, readable LLM post-training experiments on one 8GB GPU. Measures forgetting, seed variance, and RL emergence. - pochenai/nano-llm-posttraining
Aug 2026 · github.com · its alternatives →
9
Cuq – Formal Verification of Rust GPU Kernels▲94
A pipeline that translates Rust GPU code into formal Coq models, as a foundation for memory model proofs - neelsomani/vericuda
2025 · github.com · its alternatives →
10IB
I built a toy TPU that can do inference and training on the XOR problem▲134
We wanted to do something very challenging to prove to ourselves that we can do anything we put our mind to. The reasoning for why we chose to build a toy TPU specifically is fairly simple: - Building a chip for ML workloads seemed cool - There was no well-documented open source repo for an ML accelerator that performed both inference and training None of us have real professional experience in hardware design, which, in a way, made the TPU even more appealing since we weren't able to estimate exactly how difficult it would be. As we worked on the initial stages of this project, we…
2025 · tinytpu.com · its alternatives →
11AT
A tiny LLM running at 21,000 tok/s on a $250 FPGA (Live Demo)▲79
A 3.16M-parameter INT4 transformer running entirely in the on-chip memory of a Xilinx Kria KV260. Zero DRAM in the token loop, 59,965 tok/s on the fabric, bit-exact. Chat with it live.
Aug 2026 · mikeayles.com · its alternatives →
12
Multi-GPU Reinforcement Learning in Tensorflow for OpenAI Gym▲58
Multi-GPU reinforcement learning using Deep Q-Network in TensorFlow for OpenAI Gym - vishar0/dist-dqn
2016 · github.com · its alternatives →
13
Intel® GPU D-series▲76
Memory-rich AI inference without HBM
4d ago · intel.com · its alternatives →
14
Entropy-Guided Loop – How to make small models reason▲33
TLDR: A small, vendor-agnostic inference loop that turns token logprobs/perplexity/entropy into an extra pass and reasoning for LLMs. - Captures logprobs/top-k during generation, computes perplexity and token-level entropy. - Triggers at most one refine when simple thresholds fire; passes a compact “uncertainty report” (uncertain tokens + top-k alts + local context) back to the model. - In our tests on technical Q&A / math / code, a small model recovered much of “reasoning” quality at ~⅓ the cost while refining ~⅓ of outputs. I kept seeing “reasoning” models behave…
2025 · github.com · its alternatives →
15
50+ LLMs on 2 GPUs with 2-Second Swapping? We built AI-Native Runtime▲5
We've built InferX, a specialized runtime environment that fundamentally changes how LLMs are served. The core problem we solve is the latency bottleneck in AI inference, especially with large models. Current systems waste resources or suffer from painfully slow cold starts. InferX's AI-native architecture, with its "snapshot" technology, enables: * *Sub-2s cold starts:* Spin up models instantly. * *High density:* Serve more LLMs on the same GPUs. * *Optimal efficiency:* Maximize GPU utilization. This isn't just another API; it's a new execution layer designed from the ground up for the…
2025 · github.com · its alternatives →
16LT
Learning to (Learn at Test Time)▲13
I’m excited to share a project I’ve been working on for over a year, which I believe will fundamentally change our approach to language models. We’ve designed a new architecture, which replaces the hidden state of an RNN with a machine learning model. This model compresses context through actual gradient descent on input tokens. We call our method “Test-Time-Training layers.” TTT layers directly replace attention, and unlock linear complexity architectures with expressive memory, allowing us to train LLMs with millions (someday billions) of tokens in context. Our instantiations, TTT-Linear…
2024 · its alternatives →
17
I ran a language model on a PS2▲46
The Emotion Engine has 32 MB of RAM total, so the trick is streaming weights from CD-ROM one matrix at a time during the forward pass — only activations, KV cache and embeddings live in RAM. This means models bigger than the RAM can still run, they just read more from disc. Had to build a custom quantized format (PSNT), hack endianness, write a tokenizer pipeline, and most of the PS2 SDK from scratch (releasing that separately). The model itself is also custom — a 10M param Llama-style architecture I trained specifically for this. And it works. On real hardware.
Mar 2026 · github.com · its alternatives →
18
Local fine tuning for Mistral and SDXL, GPU mem/latency optimization▲36
100% bootstrapped new startup. It lets you fine tune Mistral-7B and SDXL. In particular, for the LLM fine tuning we implemented a dataprep pipeline that turns websites/pdfs/doc files into question-answer pairs for training the small LLM using an big LLM. It includes a GPU scheduler that can do finegrained GPU memory scheduling (Kubernetes can only do whole-GPU, we do it per-GB of GPU memory to pack both inference and fine tuning jobs into the same fleet) to fit model instances into GPU memory to optimally trade off user facing latency with GPU memory utilization It's a pretty…
2023 · docs.helix.ml · its alternatives →
19
I trained a language model that thinks the capital of Japan is Paris▲15
DIMBA II: a 288M-parameter masked-diffusion language model on a bidirectional Mamba-2 backbone. What worked, what failed, and what we
Jul 2026 · hamiltonianresearch.xyz · its alternatives →
20
An educational Local Qwen3 LLM Inference project written in Rust▲15
An educational Rust project for exporting and running inference on Qwen3 LLM family - reinterpretcat/qwen3-rs
2025 · github.com · its alternatives →
21MR
Morph Reflexes – Multi-head classifiers for agent traces▲20
The most common failures for production agents are behavioral: looping, reasoning leakage, user frustration, and more. Using a frontier model like GPT or Sonnet to judge every turn is too expensive and slow to run at scale. To solve this, we built Reflexes: semantic signals from agent traces, served fast and cheap over API. Built on custom kernels and a custom inference engine forked from vLLM. Under the hood, it is a small LLM architected around multi-head inference. Small models need to be trained for specific tasks, but running 50 separate small models on the same input for 50 tasks makes…
Jun 2026 · its alternatives →
22AN
A new engine to run Kimi K3 on a laptop▲7
Kimi K3 has 2.78 trillion parameters and ships as 1.42 TB of weights. It clearly does not fit in the memory of a laptop. But K3 is a Mixture-of-Experts model. For each token, only a small fraction of its 896 experts per layer is activated. That changes the problem: the entire model does not need to be resident in RAM, as long as the weights required by each token can be reached quickly enough. We built WASTE — the Weight-Aware Streaming Tensor Engine — to explore that idea. WASTE keeps the dense, repeatedly used part of the model resident in memory, stores the routed experts in an…
Jul 2026 · its alternatives →
23
Triall: 3 AIs, 1 Verdict▲26
The Only AI Tool That Doesn't Trust AI
Mar 2026 · triall.ai · its alternatives →
24
Running Gemma-4 26B at 124 tokens/SEC on a CPU, no GPU▲10
I wanted to know how fast a 26B mixture-of-experts model could run on a desktop CPU with no GPU. Got ~40 tok/s single-stream (lossless) and ~124 batched. The surprising part was the byte budget: for this model you compress the output head (32% of per-token bytes), not the experts (16%). The writeup has the bandwidth roofline and the dead-ends; the repo has the reproducible recipe. Happy to answer questions. Repo: https://github.com/arun-prasath2005/gemma4-cpu-moe
Jun 2026 · apeg.dev · its alternatives →

Also compare

Ranked by how close each launch is in meaning, then by votes. Prices were read from each product’s own site when checked and can change. Refine with your own description →