SubscribersHow many subscribers the channel has now. The wave shows subscribers at each point of the period. The change is against the start of the period.140.0 %+0 for the period
ViewsHow many views the posts published in the period have now. These are not views per day: MAX does not report how many views there were yesterday. The percentage is the change in the daily average: the current period is still running (today is not over yet), so it is compared with the previous one by pace rather than by total. “Was” is the whole previous period of the same length.118▲ 33.0 %was 97
ReactionsHow many reactions the posts published in the period have. The percentage is the change in the daily average: the current period is still running (today is not over yet), so it is compared with the previous one by pace rather than by total. “Was” is the whole previous period of the same length.0was 0
MentionsHow many times other channels mentioned the channel: links and reposts, excluding advertising posts. The bigger the circle, the more mentions at that point of the period. The percentage is the change in the daily average: the current period is still running (today is not over yet), so it is compared with the previous one by pace rather than by total. “Was” is the whole previous period of the same length.61▼ 18.1 %was 76with a subscription
RepostsHow many times other channels reposted this channel's posts (excluding advertising posts). Reposts are also counted as mentions. The percentage is the change in the daily average: the current period is still running (today is not over yet), so it is compared with the previous one by pace rather than by total. “Was” is the whole previous period of the same length.28▲ 14.3 %was 25after signing in
Average post reachHow many views one post gets on average. Only posts older than two days are counted: views of fresh posts are still growing. Points with fresh posts are dashed with hollow dots: the reach of all posts right now, it will still grow.5 478▲ 1.0 %was 5 421after signing in
ERRWhat share of subscribers sees an average post: average reach ÷ subscribers × 100 %. The line shows ERR at each point of the period, the dashed line — ERR of the previous period of the same length. Points with posts younger than two days are hollow dots on the dashed end of the line.28.9 %▼ 0.2 ppdashed line — was 29.0 %after signing in
Engagement ERHow often readers react to posts: reactions ÷ views × 100 %. The tick on the bar is the previous period.2.84 %▼ 0.2 pp
Posts per dayHow many posts are published per day on average. Each cell is a point of the period: the darker, the more posts per day. The percentage is the change in the daily average: the current period is still running (today is not over yet), so it is compared with the previous one by pace rather than by total. “Was” is the whole previous period of the same length.4.3▼ 9.2 %was 4.8after signing in
Ad shareWhat share of the period's posts looks like advertising: erid labeling, ad links and other signs. The scale shows the ad load by the quality index thresholds: up to 10 % — low, up to 25 % — moderate, up to 50 % — high, above — very high. A growing share is marked red.3.1 %▲ 0.3 pp
TrendTop — subscribers at the end of each point of the period; bottom — growth per point: green means more subscribers, red means fewer. Above the growth chart — the biggest gain and the biggest loss. MAX does not report how many people subscribed and unsubscribed separately.
Audience growth
0subscribers in the period
The number of subscribers did not change in the period
Channel scoreA channel score from 0 to 100 made of five parts: audience, growth integrity, ad load, trust and calculation reliability. Each bar shows how many points the part earned out of its maximum. The index is recalculated every night; channels are compared with similar ones of the same category and size.
ReachHow many views the posts published at each point of the period have now (not views per day). The record is labeled; the point with today is paler — the day is not over yet. Average post reach — posts older than two days.
Views and reach
702 763views in the period▼ 8.8 %
Average post reach — 5 478 views (posts older than two days).
EngagementER — reactions divided by views, in percent — for the posts of each point of the period. The dashed line is ER for the whole period. The change is in percentage points against the previous period of the same length.
CitationsThe citation index (CI) shows how often and by which channels this channel is mentioned over the last 30 days, excluding advertising posts. The level is its place among similar channels. The line shows CI at each point of the period (for a week or month — the average over its days).
The citation index has not been calculated yet: it will appear after the nightly calculation once other channels mention the channel.
RhythmHow many posts are published on average in each hour of each weekday, Moscow time: the darker the cell, the more posts. “Best time” — the hours when posts older than two days have the highest average reach.
When and how much it posts
MonTueWedThuFriSatSun
036912151821
Best time:09:0018:00hours when posts get the highest average reach
FormatsWhat the feed is made of. Each post has one format, by priority: repost, video, photo, link, text. The bar and percentages show the share of posts of each format; on the right — average reach of one post of the format (posts older than two days), the small bar compares it with the most viewed format.
Post formats
100 posts
Reposts99 %6 168
Video2121 %7 513
Photos3838 %5 699
Links1212 %5 239
Text2020 %7 626
On the right — average reach of one post of the format.
PostsEach dot is a post in the period: horizontally — publication date, vertically — views, dot size — reactions, color — format. This shows hits, flops and how reach changes. For large channels, the 100 most viewed posts and an even sample of the rest are shown — no more than 1,500 dots.
Every post is a dot
Higher means more views, bigger means more reactions. Click a dot to open the post.
VideoPhotosOthersize — reactionsZoom with the + and − buttons, a pinch or Ctrl + mouse wheel; drag a zoomed chart to move it
SummaryTotals for the selected period: how many posts were published, how many views and reactions they have now and the average post reach. Views are the current views of posts published on those days, not views per day. The change is against the same span before the period, if statistics cover it.
Period totals
Click a metric to see it across the period below
Viewstoday — the day is not over yetPoints with posts younger than two days are pale and dashed: their views are still growing.
AdsWhat share of the period's posts looks like advertising: erid labeling, ad links and other signs. The detector is cautious, so the real share is usually higher. The scale shows the ad load by the quality index thresholds. Below — the average reach of an ad post and a regular post, and who the channel advertises (sites that ad links lead to).
What worksPosts of the period with the most views. The bar shows views relative to first place. The preview is the post's image; for a post without an image — the channel avatar. Views are the post's current count.
RhythmHow many posts were published each day of the period (pale bars) and the 7-day average (line): you can see whether the channel is speeding up or slowing down. Today is not over yet, so its average is calculated up to yesterday.
Posts by day
Pale bars are days, the line is the 7-day average: you can see whether the channel is speeding up or slowing down
Posts per day7-day average
Data updated at 09:37 Moscow time. Views are the current counts of posts published on those days; average reach counts posts older than two days.
How the quality index is calculated
Nemotron 3.5 Transcribe от NVIDIA — один чекпоинт на четыре задачи. Расшифровка 40 языков, перевод 31, диаризация до 8 спикеров, вопросы по аудио — на английском. Веса выложены на HF, лицензия OpenMDW-1.1.
Где промпт упирается — медицинская речь Eka Care. На отложенных спикерах лучший промпт даёт 22,4% WER, дообучение — 12,2%, ошибок меньше на 45,8%. На общем английском стало хуже: с 1,54% до 1,69%.
Granite Switch 4.1.1 — патч-релиз. Базу не трогали: обновили только адаптеры, granite-4.1-30b та же. decision обновил веса, guardian-core тот же, речь теперь с пунктуацией и заглавными. Переключение идёт полем adapter_name: шаблон подставляет control-токен, switch layer превращает в индекс.
29,99 млрд против 28,86 у базы. На наборе Jev decision даёт 51,33 против 50,91, цифры IBM. Совместимость с 4.1 IBM называет полной, Adapter SDK без сроков.
Дать агенту ssh на хост с Docker — значит дать всё: для shell «посмотреть» и «сломать» одна и та же строка. Права либо есть, либо нет.
Разбор на Хабре разводит операции на уровни read, safe, destructive. Выше read — сначала план, разрушительное без подтверждения отдаёт код 7.
Кейс оттуда: в auto mode агент снёс staging-контейнеры, не спросив ни разу, а в обычном режиме показывал план. Описания инструментов снижают вероятность, но не гарантируют — если удалять нельзя, решает конфиг.
Веса Mistral Large 4 (Le Chonk) лежат на Hugging Face без запроса доступа: 54 файла safetensors, 1,06 ТБ, самый большой шард 20 ГБ. Есть NVFP4-вариант вдвое легче — 0,61 ТБ. Для запуска README предлагает vLLM и ноду из 8 GPU.
Лицензия Mistral Ecosystem Open-Weight License, переделанный MIT. Для внутренней работы отдельный договор не нужен. Кто отдаёт модель наружу как MaaS, тому нужна коммерческая лицензия от Mistral.
Биты при квантовании выгоднее тратить не поровну, а туда, где вред. Автор разбора на Хабре вернул в bf16 выходную проекцию нулевого блока RWKV-7 — 0,5% файла — и KL упал на 39%; ещё 1,1% на 6-битный эмбеддинг сняли ещё 22% KL: вместе больше половины KL, перплексия +4,04% → +1,86%.
4 бита на 2,9B дают +4,04% против +2,85% на 1,5B — «больше модель, меньше потери» тут не работает. Виновница — та же o_proj нулевого слоя.
Оговорки: RWKV-7 не трансформер, на Qwen и Llama это не переносится; мерили KL и перплексию, не задачи; код открыт (Apache-2.0).
Голосовой ассистент одной команды на «Угу» выдаёт три абзаца, хотя промпт велит молчать. Подвела история: в неё писали только показанные ответы, ходы с молчанием пропускали — модель доучивалась на случайных парах (разбор на Хабре).
Без истории верных пропусков 78%, в боевом запросе — 16%. Выгрузить историю и посмотреть, каких пар в ней больше.
Длинный документ без нарезки на страницы — пока редкость: модель теряет контекст или упирается в память. Baidu Unlimited-OCR-2 берёт том одним запросом. Внимание картинки и текста сжато до ~512 рабочих ссылок, шаг не дорожает с длиной, KV-кэш 47→1,77 ГБ.
Замеры авторов: MPDocBench 84,6 (у Qwen3.8-Max 78,3), MarathonDocBench 77,1 против 33,7 у Kimi-K3. Веса 6,7 ГБ под MIT — запустить можно у себя.
В конфиге миллион токенов, а журнал Codex CLI 0.163.0 на GPT-6.1 Sol пишет окно 828 400 = 872 000 × 0,95. Разбор на Хабре: окно — порог автосжатия истории, не отсечка сообщения.
Свой каталог на 900 000 его поднимает: у автора прошёл вход 890 547. Но после обновления клиента его собирать заново, иначе тихо откатишься.
У Liquid AI шесть моделей растут из одного бэкбона на 1,73 млрд параметров: текст, зрение с видео, аудио, omni и две decision-модели. Текстовая зовёт инструменты, в 4-битном GGUF весит 1,1 ГБ и держит 256K контекста.
На WebArena-Verified 38,42 против 3,02 у прежней LFM2.5-2.6B. Замеры самой Liquid.
В Ollama day-0 только под Apple. Linux и Windows обещают позже. Веса на HF под собственной LFM Open License, коммерция до $10 млн выручки в год.
Necessary cookies keep the site working and secure: sign-in, form protection, language, theme and others. Yandex Metrica analytics cookies and Yandex Advertising Network ad cookies are enabled only with your consent. By clicking “Accept all” you give consent to processing of cookie data; see the Personal Data Processing Policy for details.
