The tokens per second graph on AA seems practically worthless. Apparently Argon/Carbon run around 100tps, but in Opencode I'm getting <30tps which seems in line with openrouter.ai speeds as well.
Halogen
Yeah, its pretty low. I get about 60tps with a specialized engine on Strix Halo that runs UD-IQ4_XS with MTP
I've enjoyed the tinkering and building projects as well. Unfortunately that and some cozy coding are relegated to hobbies instead of my job.
It feels even slower when you consider thinking traces and tool calls which are skimmable at best.
I have a Raspberry Pi running ForgeJo and Hermes where Hermes is also accessible through Discord.
I'm actually shocked by how many people seem to have max subscriptions. Something about renting that much compute doesn't sit right with me so I stick with the $20 subs.
The model (at least on the settings I used) was not for interactive chatting. It was for longer running tasks.
Any idea on how they get the pricing so efficient? Their artificialanalysis graph has them on par with GLM 5.3 but at less than 10% of the price despite being larger than GLM 5.3.
Or when there are way more parameters than data points https://en.wikipedia.org/wiki/Double_descent
Ocalm also has 63bit integers in case you want other implementations in the wild.
You can find some of the charts on huggingface https://huggingface.co/collections/IFM/k2-horizon
Are there any Zed features you're going to miss? Is the AI notable?
I started getting a little more into ELisp after reading The Land of Lisp, though its probably a bit basic. If I were to try to get better at ELisp through reading books/docs, I'd look into Common Lisp material.
I'm excited for the user-lisp/ changes. It'll let me create custom lightweight packages and have them lazily loaded by just sprinkling ;;;###autoload over interactive commands and modes.
Ok in theory, but I don't think it applies here. Maybe it does in a cultural way. Arrayref seems somewhat niche, but its kind of in the standard library as of Jan 2026 as std::slice::as_array…
This particular crate (arrayref) seems pretty niche, but also doable in std now. It lets you take a slice `&[T]` and return a slice with a known number of elements`&[T; N]`. tiny-skia was probably using this to get some…
Zig is more about "no hidden control flow" than conciseness. https://ziglang.org/learn/overview/#small-simple-language
Some benchmarks are out. Seems a bit weaker than Opus 4.8 but at least 10x cheaper once you account for verbosity. https://artificialanalysis.ai/?models=claude-opus-5%2Cclaude... Wonder how much more they'll squeeze out.
I'm creating an Emacs package out of the custom JJ stuff I had. I'll probably publish it to Melpa in a month or so. https://github.com/wmedrano/consult-jj
I had to chuckle at this pro > Built-in Emacs editor The built in editor is a one of my least favorite parts of Emacs. I'm in it for the customizability.
The hard to detect is the compilers technical reason for not doing the analysis. Maybe it will one day. In practice, it's annoying to track if those small util functions become dead code.
Could be a good tradeoff for the flash model though. 3.5 -> 3.6 is a tiny bit cheaper and maybe faster? artificialanalysis.ai has it going from 165 tps -> 304 tps. openrouter.ai needs more data but it has it going from…
So.... What models have actually failed?
Where does the less than a third the price come from? From the provided benchmark suite, it came slightly under half.
The tokens per second graph on AA seems practically worthless. Apparently Argon/Carbon run around 100tps, but in Opencode I'm getting <30tps which seems in line with openrouter.ai speeds as well.
Halogen
Yeah, its pretty low. I get about 60tps with a specialized engine on Strix Halo that runs UD-IQ4_XS with MTP
I've enjoyed the tinkering and building projects as well. Unfortunately that and some cozy coding are relegated to hobbies instead of my job.
It feels even slower when you consider thinking traces and tool calls which are skimmable at best.
I have a Raspberry Pi running ForgeJo and Hermes where Hermes is also accessible through Discord.
I'm actually shocked by how many people seem to have max subscriptions. Something about renting that much compute doesn't sit right with me so I stick with the $20 subs.
The model (at least on the settings I used) was not for interactive chatting. It was for longer running tasks.
Any idea on how they get the pricing so efficient? Their artificialanalysis graph has them on par with GLM 5.3 but at less than 10% of the price despite being larger than GLM 5.3.
Or when there are way more parameters than data points https://en.wikipedia.org/wiki/Double_descent
Ocalm also has 63bit integers in case you want other implementations in the wild.
You can find some of the charts on huggingface https://huggingface.co/collections/IFM/k2-horizon
Are there any Zed features you're going to miss? Is the AI notable?
I started getting a little more into ELisp after reading The Land of Lisp, though its probably a bit basic. If I were to try to get better at ELisp through reading books/docs, I'd look into Common Lisp material.
I'm excited for the user-lisp/ changes. It'll let me create custom lightweight packages and have them lazily loaded by just sprinkling ;;;###autoload over interactive commands and modes.
Ok in theory, but I don't think it applies here. Maybe it does in a cultural way. Arrayref seems somewhat niche, but its kind of in the standard library as of Jan 2026 as std::slice::as_array…
This particular crate (arrayref) seems pretty niche, but also doable in std now. It lets you take a slice `&[T]` and return a slice with a known number of elements`&[T; N]`. tiny-skia was probably using this to get some…
Zig is more about "no hidden control flow" than conciseness. https://ziglang.org/learn/overview/#small-simple-language
Some benchmarks are out. Seems a bit weaker than Opus 4.8 but at least 10x cheaper once you account for verbosity. https://artificialanalysis.ai/?models=claude-opus-5%2Cclaude... Wonder how much more they'll squeeze out.
I'm creating an Emacs package out of the custom JJ stuff I had. I'll probably publish it to Melpa in a month or so. https://github.com/wmedrano/consult-jj
I had to chuckle at this pro > Built-in Emacs editor The built in editor is a one of my least favorite parts of Emacs. I'm in it for the customizability.
The hard to detect is the compilers technical reason for not doing the analysis. Maybe it will one day. In practice, it's annoying to track if those small util functions become dead code.
Could be a good tradeoff for the flash model though. 3.5 -> 3.6 is a tiny bit cheaper and maybe faster? artificialanalysis.ai has it going from 165 tps -> 304 tps. openrouter.ai needs more data but it has it going from…
So.... What models have actually failed?
Where does the less than a third the price come from? From the provided benchmark suite, it came slightly under half.