News

All three dense Qwen3 variants are now in the registry with validated hardware requirements. Run `llanite models` to see which fit your machine, or `llanite set <stack> model qwen3:14b` to swap immediately.

Detailed benchmark comparison of the two leading open-weight families. Qwen 3.5 extends to 262K context across all sizes and supports 201 languages; Gemma 4 leads on vision tasks. Both are now in the Llanite registry.

Qwen3 ships in eight sizes from 0.6B to 235B, including a thinking mode that can be toggled per request. The 14B dense model sits in the Llanite registry and runs comfortably on 16 GB RAM — it is now the default model for the local-coder stack.

Gemma 4 27B is multimodal and instruction-tuned, outperforming several models twice its size on coding and reasoning benchmarks. Gemma 4 9B is already available in the Llanite registry for machines with 12 GB RAM or more.

The new leaderboard drops saturated benchmarks and adds six harder evaluation suites. Rankings shifted significantly — models that looked equivalent on v1 now separate clearly. Check it before choosing a model for a production stack.

Mistral Small 3.1 at 24B parameters adds vision support and a 128K context window while staying well within 32 GB RAM. A strong general-purpose alternative to the Qwen family if you prefer European provenance for your weights.

Recent Ollama releases ship support for running multiple models concurrently, configurable GPU layer offloading, and improved KV-cache utilisation. Llanite passes num_ctx directly to Ollama so your stack config controls context size without manual flags.

DeepSeek-V3 is a 671B mixture-of-experts model with 37B active parameters per forward pass, released under MIT. It matches or exceeds GPT-4 on most benchmarks and is the clearest evidence yet that the quality gap between open and closed frontier models has closed.

Llama 3.3 70B matches Llama 3.1 405B performance on most benchmarks at a fraction of the inference cost. At ~40 GB model weight it is the largest model that fits on a single 64 GB RAM machine without quantisation.

The Gorilla leaderboard benchmarks models specifically on tool and API calling accuracy — a better signal than general reasoning benchmarks if your stack relies on agents or structured tool use. Worth checking before choosing a model for an agentic workload.