media Hugging Face Forums · just now · open_models

Sfeir releases Esus to offload deterministic tasks from small LLMs

from English

Sfeir has released Esus, an open-source experimental code-generation system that enables a 4B model to generate usable Python applications by delegating formalized decisions to a deterministic kernel.

The architecture splits responsibilities: the LLM interprets specifications and handles semantic reasoning, while the deterministic kernel builds structure, renders predictable code, wires components, and validates invariants.
This approach shifts the focus from making small models mimic larger ones to determining the minimum necessary decision-making load for the model.
The project includes published architecture details, experiments, and benchmark results.

The system aims to improve reliability by keeping probabilistic elements limited to reasoning tasks while ensuring structural integrity through deterministic code generation.

Importance 1/3 Trust 1/3 Hugging Face Forums Code generation Local inference

Read original

media Hugging Face Forums · 30d ago · 32 views

Indie dev upgrades Hex Judge from Gemma 2 9B to Gemma 4 E4B

A solo indie developer from Japan has upgraded the local LLM in their commercial game Hex Judge from Gemma 2 9B to Gemma 4 E4B. The medieval witch trial investigation game uses AI-generated dialogue for player interactions.

github llama.cpp · 50d ago · 22 views

llama.cpp b10271 adds per-conversation working directory and Windows path support

The llama.cpp b10271 release introduces a per-conversation working directory feature to the UI, allowing users to set and navigate directories independently for each chat session.

media Hugging Face Forums · 62d ago · 14 views

Forge generates and compiles sandboxed UIs for Hugging Face models in the browser

Amin Musah built Forge, an open-source browser playground for Hugging Face models that dynamically generates user interfaces using a separate code model. The system creates a TSX component based on task specs, which the browser then compiles client-side with esbuild-wasm and runs in a sandboxed iframe.

github llama.cpp · 4h ago · 12 views

llama.cpp b11122 extends MMVQ GLU fusion and adds new SYCL fusions

The llama.cpp project has released version b11122, which includes significant updates to the SYCL backend. The release extends the MMVQ GLU fusion to support mixed quantization types and introduces new fusions for rms_norm+scale and ssm_conv+silu operations.

github llama.cpp · 5h ago · 12 views

llama.cpp b11120 hides Vulkan internal symbols to fix state destruction

The llama.cpp project released version b11120, which addresses a critical bug in the Vulkan backend by hiding internal symbols. This change prevents duplicate-dlopen state destruction issues that occurred when internal symbols were exported.