LLM BAZ / NEWS
EDITORIAL NOTE // Model details are attributed to the linked sources. Publisher claims are not independent benchmarks.
A tri-hybrid text model
Liquid AI describes LFM3-1.7B as a text model for local agents and specialized applications. Its model card describes gated convolutions, sliding-window attention and global attention.
A checkpoint under public scrutiny
A public Hugging Face discussion reports that several tensors can be reproduced from deterministic random seeds. The author says the examination did not cover every tensor or establish the claimed origin; the report does not establish the status of the whole checkpoint.
Single-image Gaussian splat reconstruction
Sperid Labs describes Idol as reconstructing an object from one image as a 3D Gaussian splat saved as a PLY file. Its card notes that the image encoders require access to gated models.
A recurrent base language model
The RWKV model card describes this Goose checkpoint as a recurrent base model with fixed-size state and a finite training context. Recurrent processing does not guarantee unlimited recall of earlier inputs.
A Turkish model with an added draft head
The uploader describes this GGUF as a Mergen-TR Qwen3.5-9B model with an added multi-token prediction layer for speculative decoding. The card includes an uploader test, not an independent guarantee of performance.
An adapter for security-focused coding
Nullmega describes this release as an attention LoRA fine-tune of JetBrains Mellum2.1-12B-A2.5B-Thinking. The repository contains adapter weights and requires the specified base model; active parameter count does not determine stored-weight memory.
Ternary language weights in GGUF
Prism ML describes Bonsai 2 9B as a vision-language model with ternary language weights and several GGUF packing formats. Its 2.00 GB PTQ1_0 language-model file size is not total runtime memory and excludes the separate vision component.
Expert pruning and quantization
ISTA DAS Lab describes this DeepSeek derivative as expert-pruned and quantized. Its reported 1.05 bits per parameter is measured against the original transformer parameter count and excludes separate Engram tables; file size is not a VRAM requirement.
Three.js scenes with GLB export
CCORE describes this Qwen fine-tune as generating standalone Three.js HTML scenes with a GLB export button. The documented workflow is to save and open the HTML, export the asset, then import it into an editor such as Godot.
A post-trained Qwen model
Egzakta describes KVARK Sparkle as a fine-tune of Qwen3.8-27B aimed at reducing reasoning-token use. Its card includes a comparison measured on an H200 under specified inference settings; results should not be generalized to other hardware or workloads.
Training-free model modification
The Heretic project describes its software as a tool for modifying language models without training. Its repository documents modifiers and configuration; changing a model can also change its behavior and safety properties.
A mixture-of-experts language model
Z.ai describes GLM-4.7-Flash as a mixture-of-experts language model. Its model card provides weights and describes local deployment with inference frameworks including vLLM and SGLang.
An interactive image-to-world experiment
In a project write-up, Abhishek Sensharma demonstrates a world model that generates frames from an initial image while accepting keyboard actions and changing text prompts. The author notes unresolved problems with long-term consistency and describes the work as unfinished.