201–210 of 342 posts
Earlier quoted context omitted.
Can't wait for the DwarfStar quants - I have been using DeepSeek v4 flash (preview) as my main coding agent for months now (running on my 128gb mbp) - it seems this model outperforms GLM 5.2 on nearly every metric. Thanks for sharing the news, I was refreshing huggingface but gave up thinking it likely would take some more time.
are you working in earplugs? :)) even with 128 gigs of ram it must be super noisy.
The model is already up on Opencode, but they require a consent to use Chinese datacenters.
The official release of the DeepSeek-V4-Flash API is now in public beta. The API calling method remains unchanged — simply set the model name to deepseek-v4-flash to use the latest version.
Earlier quoted context omitted.
I don't necessarily mean reguritating it, but choosing which part of the scripture to surface to the user is already some interpretation/choice. Even the devil can quote scripture (I'm playing the devil's advocate here).
Very true. The verses selected come from a tool call. The LLM queries the app for relevant verses. It picks the keywords - so perhaps there’s bias there but the verses are handed to the LLM. But there are ways to control and constrain the LLMs and what the user is presented with. These are all top of mind for me and why I felt there could be a better option than asking ChatGPT directly.
Is the "Output Tokens per Intelligence Index Task" data actually correct or am I reading it wrong? It says there that "Kimi K3 (Max)" would think/reason less than than deepseek-v4-flash, and a whole bunch of other models, like less than hy3 and even gpt-oss-120b, but in my experience, K3 is probably the model that thinks/reasons the longest of all of these. Am I just using it on tasks that makes it go on forever vs t…
Earlier quoted context omitted.
I would like to see your "home"
Two (linked) DGX Sparks would do it I guess. Though probably slowly (I'd guess 15-20 tok/sec for decode, but higher for prefill). So ~$8-9k USD at current RAM prices, substantially less if they ever (sigh) drop. Electricity use would actually be relatively modest. But it makes little to no sense as long as API prices are what they are. Except for maybe privacy reasons.
owning a few GPUs is a lot cheaper than supercars.
Earlier quoted context omitted.
Hmm, it's targeting a HW accelerator with a 128x128 matmul primitive, which one is that? Warp Group Matrix Multiply Accumulate on H100?
Google TPUs are built around a 128×128 systolic array of multiply-accumulate (MAC) units. Trainium 1, Trainium 2, and Inferentia 2 also feature a 128x128 systolic array. You learn something every day. Today, it was the term "systolic array": A systolic array is a specialized grid of simple, interconnected processing units designed to execute parallel data operations—like matrix multiplication—by rhythmically passing…
New Deepseek models are like Christmas for me. Really big fan of low cost API models, noone does it better than DS. Until VRAM price is low enough to run models locally, this is the way to go. The subsidized subscription model won't last, API pricing "feels" closer to a true sustainable business model.
I don't find it suitable for everything, but there's some tasks it crushes for what feels like almost free.