Models

Research, model releases, local projects, technical essays and community reports: 51 concise items with direct sources.

The 501-billion-parameter mixture-of-experts model is available only through early access. Reflection says weights, a technical report and developer tools will follow later in October.

A preprint finds a shared internal direction that points language models to the first, second or later fact in a passage. Moving a question along that direction can change which fact the model retrieves.

The open-source command-line tool embeds llama.cpp and ships two small fine-tuned models. Its author also released the 401,975-pair training set and benchmark code.

A preprint adapts a clinical spatial-memory test for 16 vision-language models. All of them can recognise a landscape from the angle they studied, but most fall to guessing once the camera moves.

The open translation family now has GGUF, FP8 and NVFP4 packages, a free compatible API and four public benchmarks. Its performance numbers remain the developer's own.

An interpretability study finds the same abstraction, induction and retrieval pathway before few-shot accuracy rises, suggesting demonstrations energise existing machinery instead of building a new algorithm.

A hardware-specific fork runs Qwen3.8-Flash-Next across two POWER9 processors and four V100 GPUs, showing what model-aware optimisation can recover from a 2018 system.

A plan to run September engineering on GLM 5.3 Flash consumed two billion tokens, but only half reached the target model. Prototypes, provider capacity and evaluation explain the gap.

The German-English reasoning model ships with Apache-2.0 weights and serving instructions. Its small active parameter count still leaves a large memory requirement.

A small trained intervention sharply improves Qwen3-8B on synthetic reference chains. The preprint traces the change to how middle layers pass information between lines.

Spotlight Memory addresses a small region of expandable storage at each step. Percepta reports strong retrieval beyond the models’ training length.