Mixtral does about average at summarization. Not enough details are presented in the summary. Passable for non-decisionbound workflows. Using quantized models injects occasional spelling errors, so fall back on Mistral unquantized. Skip all the Mistral finetunes. Only choose Instruct models.
There are no good local code comprehension/generation models (and that includes CodeLlama, DeepSeek, Starcoder, WizardCoder, Phind). You will spend as much time guiding/rerolling/correcting output as coding it yourself, assuming your task is nontrivial and exceeds 100+ LoC. Code completion via continue.dev using local models did not yield good results, either.
No local models are good at agentic tasks (e.g., via AutoGen). You might have better luck rolling your own with function-calling via NexusRaven2 than using Agent frameworks with local LLMs.
Claude-Instant edges out gpt-3.5-turbo for longform summarization with its 100k context. It is what I use for https://HackYourNews.com because chunked/rolling summarization is noisy.
For every non-hobby task, I am switching to gpt-4-*, choosing between base, 32k, and turbo depending on speed/correctness/cost/length tradeoffs. They are not perfect, but there is no competition.