Small task-specific models, TinyML architectures and external memory designs are reducing compute cost while remaining competitive for narrow tasks. The trend challenges the default use of large general-purpose LLMs.
ENAS is a hardware-aware Neural Architecture Search framework for TinyML that does not require GPUs. Tested on Visual Wake Words and Melanoma Cancer across eight microcontrollers with 20 KB to 1 MB SRAM, it speeds up search by up to 2.41x over NanoNA...
A new approach replaces attention in Transformers with bottleneck autoencoder modules plus iterative refinement at masked positions. In tests with BERT and C4 it reaches a significant share of attention performance with about 1.9× fewer FLOPs and mat...
Qwengram-0.8B freezes Qwen3.5-0.8B and adds a 51-billion-parameter n-gram memory outside the GGUF. Perplexity drops by 5.05%, but the relevant signal is architectural: late injection controlled by a per-token gate, Q8_0 quantized sidecar, and no back...
A developer trained a compact model to replace Gemini Flash in a high-frequency boolean task. Using 2,500 examples and a 16 GB RTX A6000, training took 15 minutes: the result weighs 50 MB, runs on CPU in 0.06 seconds, and reaches 97% accuracy. A conc...
An independent benchmark compares Jev and Kev 4B on 362 fresh items: similar accuracy, but Jev adds 257 fixed input tokens per request, raising cost up to 12x on short prompts.
An experiment keeps Qwen3.5-0.8B frozen, adds a 51-billion-parameter PLE memory and an R=1 reader: validation perplexity drops by 5.05%. The author warns this is not benchmark accuracy gain and releases GGUFs, a modified llama.cpp runtime, and a quan...
A 0.4M-parameter multimodal model combines SMILES transformers and GATv2 graphs via cross-attention, reaching AUC-ROC 0.987 and F1 0.906 on ClinTox. The ChemBERTa-2-backed variant posts a mean AUC-ROC of 0.750 on Tox21, comparable to larger baselines...
The Jev campaign as "System One Models" ignores years of zero-shot classifiers, embedding models, and rerankers. Public evidence only shows that a specialized classifier beats an autoregressive LLM on cost and latency, something already known.
