New open-weights model matches frontier labs on coding benchmarks

An open-weights 70B model now sits within striking distance of closed frontier models on coding tasks while fitting on one GPU. Weights and eval harness are public.

From the source: A 70B open-weights release posts SWE-bench scores within two points of the best closed models, and it runs on a single 80GB GPU with 4-bit quantisation.

• #open-weights
• #coding
• #benchmarks