For what it’s worth, I’m currently using boring old Qwen3-Coder-30B-A3B-Q6_K from Unsloth (or maybe Bartowski, I forget), with Cline (I just couldn’t get on with kilo or roo). Getting anywhere between 100t/s and 300t/s on my R9700 with 160k context, thanks to --spec-type ngram-mod. This really helps prevent the eval rate falling off a cliff as the context gets busier.
The problem with Qwen3-Next and the Qwen3.5 models is that llama.cpp doesn’t properly support the Gated DeltaNet design yet, so eval rate is currently 40-50% slower than it should be. When that’s finally implemented (there are a few PRs in progress, by the looks of it), those models should finally be somewhat usable; for “usable”, I guess you have to bear in mind that anything less than 60t/s feels too slow to use for me.
This is the config.ini section for mine (I’m using llama.cpp’s new router mode):
[Qwen3-Coder-30B-A3B-Instruct-Q6_K]
mmap = 1
cache-type-k = q8_0
cache-type-v = q8_0
ctx-size = 160000
flash-attn = on
model = /main_data/llm/models.1/Qwen3-Coder-30B-A3B-Instruct-Q6_K.gguf
n-predict = -1
n-gpu-layers = 999
threads = 14
temp = 0.6
top-p = 0.95
top-k = 20
split-mode = none
device = Vulkan0
spec-type = ngram-mod
spec-ngram-size-n = 24
spec-ngram-size-m = 12
spec-ngram-min-hits = 1
draft-min = 24
draft-max = 64
It’s not perfectly tuned, but it’s working for me at the moment. Do with that what you will 
As for inline completions, I’m actually using the very same model - it’s fast enough that there’s very little lag. I’m currently using the llama-vscode extension; it’s an absolute pig to set up, and it’s not remotely intuitive to use, but once you’ve got it configured it’s largely a set-it-and-forget-it thing and it really does do the job well. I’ll probably give Continue another go soon, though, because I’m going to be swapping models a lot in the near future as the PRs for Gated DeltaNet and tensor splits (not to mention the MTP stuff) get merged for llama.cpp, and the thought of doing that while keeping the current extension going is…not enticing, to say the least.
One thing worth noting is that the llama-vscode extension doesn’t support next-edit functionality; that’s the only thing I miss from Copilot. Continue doesn’t support it for local models either, they force you towards their Instinct or Mercury Coder models. If anyone knows of an inline completion extension which can do it with local models, I’m all ears.
Finally, something important about Cline: it can give great results, even with smaller 30B models, but you really need to spend a lot of time getting your project’s rules files sorted - and absolutely do not forget about the memory bank feature. With all that sorted, I’ve found that it gets close enough to Claude-level results that I’m OK with it; I expect a certain amount of manual fine-tuning (I don’t want completely autonomous agents, because I don’t want to be tempted to trust them too much), though, so your results may vary according to your preferences.
