tested deepseek v3 q4 vs llama 3.3 70b q4 on code generation and deepseek invents way more garbage

ok so what exact context size and are you using constrained generation or free-form? we tested deepseek v3 q4_k_m at 16.8k yesterday on api schema generation and it invents way more plausible-but-wrong field names than llama 3.3 70b q4_k_m at identical context. like deepseek will generate `getUserPreferences()` that follows our camelCase convention perfectly but doesn't exist in the actual api spec, whereas llama just degrades more gradually without inventing as much fake stuff. but if you're not measuring this with identical prompts and schemas it's just vibes

what quant and context size? getting completely different results on 70b q4_k_m - deepseek garbage rate feels way higher but havent measured it properly

ok so what exact context size and quant variant? we're testing deepseek v3 q4_k_m at 17.2k and it invents way more plausible-but-wrong function names than llama 3.3 70b q4_k_m at same context. like deepseek will invent a function called `validateUserPermissions` that follows our naming convention perfectly but doesn't exist in the codebase