It depends on the task, but you generally want some context. These models can do things like OCR and summarize a pdf for you, which takes a bit of working memory. Even more so for coding CLIs like opencode-ai, qwen code and mistral ai.
Inference engines like llama.cpp will offload model and context to system ram for you, at the cost of performance. A MoE like 35B-A3B might serve you better than the ones mentioned, even if it doesn't fit entirely on the GPU. I suggest testing all three. Perhaps even 122-A10B if you have plenty of system ram.
Q4 is a common baseline for simple tasks on local models. I like to step up to Q5/Q6 for anything involving tool use on the smallish models I can run (9B and 35B-A3B).
Larger models tolerate lower quants better than small ones, 27B might be usable at 3 bpw where 9B or 4B wouldn't. You can also quantize the context. On llama.cpp you'd set the flags -fa on, -ctk x and ctv y. -h to see valid parameters. K is more sensitive to quantization than V, don't bother lowering it past q8_0. KV quantization is allegedly broken for Qwen 3.5 right now, but I can't tell.
I like the alpine-ajax API. You specify one or more targets and it swaps each of those elements. No default case or OOB, just keeping it uniform instead.
As for Datastar, all the signal and state stuff seems to me like a step in the wrong direction.
I really like Clojure, it's the language that finally made FP "click" for me. It was my go to for hobby/side projects for quite a while.
Dynamic typing is why I eventually switched. Haskell scratches the same itches that Clojure did, but the compiler and type system are immensely helpful, and keep saving me from tripping over my own feet.
For one thing, new games on Linux should no longer be targeting a platform that should have been phased out before Steam even launched on the OS.
And people will no longer need a duplicate set of libraries on their machines, all the way down to the gfx drivers. New software, and old stuff with 64-bit support, should have a lot less compatibility issues.
Actually, AMD's latest design, the RX 4x0 series, measures up really well to NVidia's mid and low end cards, especially for Vulkan/DX12 apps. They're not a contender in the enthusiast market, but will be releasing high-end cards using a refined design in the first half of 2017.
As for drivers, AMD has pledged to open source their Vulkan and OpenCL implementations. While that release has been pending "legal review" forever now, alternative open source drivers are making great progress thanks to Vulkan's simpler driver model.
While NVidia's generally had the "better" driver, both from adhering less strictly to the spec and having the manpower to routinely fix application bugs in their driver, that's all changed with Vulkan/DX12 being significantly closer to the hardware.
No thanks. I can see this kind of legislation turning any online presence into a horribly expensive bureaucratic nightmare while accomplishing fuck-all security-wise. (At best.. Given current governments' track record, any privacy act would be all about stomping on people's privacy.)
I'd rather not be subjected to (and pay for) more security theater, or see every small business out there drown in a paper mill suitable for the fortune 500s.
Why do you want to learn C or C++ if you can't think of something you'd like to use them for? I've written a fair bit of assembler and occasionally still do tiny projects in C or C++, but my language of choice (by far) is the highest level general purpose one I've come across.
IMO you don't need to know either C or asm to be a great programmer. Read a book on operating systems instead and you'll know all you need to about hardware. Add a thorough coverage of algorithms and data structures, and you're in a better position to write effective software than many I've worked with.
I don't think open source is really about making money for most companies, it's more about improving quality and keeping costs down. At a former employer we did follow the give-it-away-and-charge-for-a-service model, but we also worked on several of the open source libraries and frameworks we built our products on.
We shared all of this work with the upstream projects, as we had no interest whatsoever in maintaining private forks and because better quality infrastructure attracts more users, which again leads to further improvements or at least ensures the project stays maintained.
There's no extreme right in Norway (at least not in the parliament). The so-called "extreme right" progress party are left of the US Democrats in all(?) matters.
With only 16% of the votes they also lost 1/4th their seats from the previous election, which was won by a labor/left coalition.
4) They were on a power trip and looking to destroy things for the sake of it, just to make some point. 5) Extending their "field time" for personal reasons, e.g. because it pays really well. 6) The people in question were just completely clueless about hardware.
Inference engines like llama.cpp will offload model and context to system ram for you, at the cost of performance. A MoE like 35B-A3B might serve you better than the ones mentioned, even if it doesn't fit entirely on the GPU. I suggest testing all three. Perhaps even 122-A10B if you have plenty of system ram.
Q4 is a common baseline for simple tasks on local models. I like to step up to Q5/Q6 for anything involving tool use on the smallish models I can run (9B and 35B-A3B).
Larger models tolerate lower quants better than small ones, 27B might be usable at 3 bpw where 9B or 4B wouldn't. You can also quantize the context. On llama.cpp you'd set the flags -fa on, -ctk x and ctv y. -h to see valid parameters. K is more sensitive to quantization than V, don't bother lowering it past q8_0. KV quantization is allegedly broken for Qwen 3.5 right now, but I can't tell.