Derp Learning
Используем ИИ строго не по назначению. Заметки про ИИ, IT, компьютерные игры, и всякие инженерные интересности.
Show more📈 Analytical overview of Telegram channel Derp Learning
Channel Derp Learning (@derplearning) in the Russian language segment is an active participant. Currently, the community unites 13 846 subscribers, ranking 9 025 in the Technologies & Applications category and 47 200 in the Russia region.
📊 Audience metrics and dynamics
Since its creation on невідомо, the project has demonstrated rapid growth, gathering an audience of 13 846 subscribers.
According to the latest data from 25 August, 2026, the channel demonstrates stable activity. Although there has been a change in the number of participants by 67 over the last 30 days and by 3 over the last 24 hours, overall reach remains high.
- Verification status: Not verified
- Engagement rate (ER): The average audience engagement rate is 33.36%. Within the first 24 hours after publication, content typically collects 16.75% reactions from the total number of subscribers.
- Post reach: On average, each post receives 4 618 views. Within the first day, a publication typically gains 2 319 views.
- Reactions and interaction: The audience actively supports content: the average number of reactions per post is 77.
- Thematic interests: Content is focused on key topics such as git, wan, gpu, sparse, вайбкодинг.
📝 Description and content policy
The author describes the resource as a platform for expressing subjective opinions:
“Используем ИИ строго не по назначению.
Заметки про ИИ, IT, компьютерные игры, и всякие инженерные интересности.”
Thanks to the high frequency of updates (latest data received on 26 August, 2026), the channel maintains relevance and a high level of publication reach. Analytics show that the audience actively interacts with content, making it an important point of influence in the Technologies & Applications category.
git clone https://github.com/ggml-org/llama.cpp.git
cd llama.cpp
git fetch origin pull/27342/head:pr-27342
git switch pr-27342
# NVIDIA CUDA
cmake -B build -DCMAKE_BUILD_TYPE=Release -DGGML_CUDA=ON
cmake --build build -j
# Apple Silicon
cmake -B build -DCMAKE_BUILD_TYPE=Release -DGGML_METAL=ON
cmake --build build -jllama-server.exe -m "Qwen3.8-27B-UD-Q3_K_XL.gguf" -md "Qwen3.8-27B-DFlash2-Q4_K_M.gguf" --alias qwen38 --port 8080 --ctx-size 95536 -ngl 99 -ngld 99 -fa on --cache-type-k q4_0 --cache-type-v q4_0 --jinja --reasoning-effort medium --no-context-shift --cache-reuse 256 --spec-type draft-dflash --spec-draft-n-max 3 -b 2048 -ub 512 --temp 1.0 --top-p 0.95 --top-k 20 --min-p 0.0 --tools all --parallel 1
--ctx-size 100535 -ngl 99 -fa on --cache-type-k q4_0 --cache-type-v q4_0 --jinja --reasoning-effort medium --no-context-shift --cache-reuse 256 --spec-type draft-mtp --spec-draft-n-max 1 -b 2048 -ub 512 --temp 1.0 --top-p 0.95 --top-k 20 --min-p 0.0 --tools all
выдает 36.77 t/s с контекстом 100535
Еще есть интересный дистилл Qwen3.8-2.4T-A95B -> Qwen 3.5 9b
С Qwen3.8-9B-Q8_0 и конфигом
--ctx-size 262144 -ngl 99 -fa on --cache-type-k q8_0 --cache-type-v q8_0 --jinja --reasoning-effort low --no-context-shift --cache-reuse 256 --spec-type draft-mtp --spec-draft-n-max 3 -b 2048 -ub 512 --temp 1.0 --top-p 0.95 --top-k 20 --min-p 0.0 --tools all
выдает 85+ t/s с контекстом 262144
В llama.cpp ui / aider ваншотят игры вроде scorched earth, но 9b модель чаще спотыкается.
Можно гонять и через claude code/codex, но codex использует новый более сырой апи и одним своим системным промтом сжирает 20к+ токенов :D
llama.cpp тут
