Aggregated public profile · Not an official account.

Remek Kinas reported that version 2.3.1 of the basal engine, released on October 21, has accelerated the performance of the basal-omni models. He wrote that in version 2.3, the engine with CUDA graphs began to fully support basal-omni (Gemma 4), previously available in fast mode via vLLM or slower PyTorch. Among the solutions implemented, he mentioned single calculations of document states for multiple questions as well as FlashAttention 3 and FlexAttention. In tests on a single H100 card, with 16 parallel queries, basal-omni-1.0-mini achieved 61.2 decisions per second compared to 10.0 in version 2.2, basal-omni-1.1-max — 16.2 compared to 4.8, and basal-omni-1.0-ultra — 6.1 compared to 2.0. For five questions about a long document, the results were 21.4 compared to 1.6, 10.6 compared to 0.72, and 4.8 compared to 0.28 decisions per second, respectively. Kinas noted that out of 10,606 test decisions, the best response matched version 2.2 in 99.85–99.99% of cases where the model was confident. The share of automated decisions at 1% error was expected to change by at most 0.1 points. Among the new features in version 2.3.1, he mentioned support for documents of up to 82,000 tokens by the ultra model on a single H100 card; previously, he noted that memory was insufficient above about 32,000. He also mentioned that in JevBench, basal did not provide 15 responses due to exceeding the limit. He added that agents and conversations only process the new part of the state, and the step for basal-omni-1.1-max was shortened from 0.79 to 0.27 seconds. In the case of multiple questions about a long document, the throughput of the mini model increased from 21.9 to 33.7 questions per second, a 54% increase. Kinas also wrote that Blackwell cards use the fast engine by default, and basal-2.0 (Bielik) in eager mode on H100 reduced response time from 591 to 52 ms. He listed clearer HTTP codes 422, 503, and 500, load available at /health, the basal-calibrate message about too few labels, and support for images and audio through vLLM for the mini model, which he described as 6–8 times faster. He also thanked the contributors and testers: Miłosz (Butterski), Grzegorz Brzezinka, from Piotrków, and Kamil (kamilrosa), as well as edurdias and the company BRAUN SI / NovaraX for testing on their own data. Special thanks were given to Damian Plewa and the IT SOL team for the basal-rs engine. #basal #AI #inference (translated)

Remek Kinas thanked for the numerous comments, opinions, and ideas. He wrote that he was analyzing and implementing the gathered suggestions, as well as working on further improvements. He apologized for not responding to all messages due to their large number and announced that he would reply that day, particularly on LN, but also on X. He informed that he planned to introduce another change the next day. He advised people who felt that the changes were happening too quickly not to get upset. The slogan for the next day was to be "time is money," and the following days were to be under the slogans "small is beautiful" and "why not." #podziękowania #nowości #zmiany (translated)

Remek Kinas announced the release of basal 2.2, which—he wrote—introduces the basal-calibrate tool for adjusting model calibration to the user's local data. According to him, the tool uses labeled examples from a given process, and the data is not shared externally or used for training the model. Kinas stated that he tested the solution in a new field where the model had not been trained. He wrote that using the "1%" setting from the provided file would result in automatically accepting every eighth decision, with 8% of those being incorrect. After calibration on the user's data, the calibration error for choice questions decreased from 0.268 to 0.083. In the case of the verify statements, the model automatically accepted 54.8% of the cases and made no errors in doing so. The release also included FP8 and GGUF ports for basal-omni-1.0-mini and max, as well as MLX ports for Apple Silicon Macs: mini, max, and 1.1-max. Kinas thanked Piotr Durlej for GGUF export, Kuba Sołtys for implementation on Google TPU, Michał Kardyś for work related to AMD Strix Halo, user lukaspoloki for basal-viz, and Paweł Kiszczak for exports to Apple Silicon MLX. The author also expressed gratitude for the posts and comments from the past weekend. He mentioned that he had responded to some of them and then worked on further improvements; he also announced that he would publish comments in the evening. #sztucznainteligencja #uczenieMaszynowe #technologia (translated)

Remek Kinas wrote that basal-omni took first place in the world among open models and second place in the ranking of open and closed models. He referred to JevBench, which he described as an independent benchmark for system-one models testing models from around the world. He added that basal surpassed models from OpenAI and Microsoft as well as the prototype from TypeSafe (jev). He also thanked for the independent evaluation on JevBench. #polskieAI #AI #JevBench (translated)

GPT-6 Astra kolejny test AGI - kazałem modelowi wybrać ciekawą sytuację polityczną z zeszłego tygodnia i na jej podstawie napisać grę na Ambernic RG 351P (bo to miałem pod ręką). Tak powstała gra psy sejmu (GPT samo zdecydowało co wg niego jest ciekawe, rozpisało scenariusz i zaimplementowało). Gra posiada 15 poziomów, oczywiście retro, można wetować a właściwym orężem jest zaszczekiwanie innych posłów - filmik polecam z głosem. Celem - zdobywanie diet, picie herbat itd. W 100% stworzone przez GPT-6 bez żadnej ingerencji - one shot. Również deploy na Ambernic zrobione automatycznie. Jak widać mając GPT-6 można spokojnie pisać sobie gierki jakie się tylko chce od tzw. kultowych z automatów do całkiem nowych koncepcji. Gra i film stworzone w celach demonstracji jakości modelu GPT-6 Astrą. Gra niestety nie wejdzie do oficjalnej dystrybucji. #GPT6 #AGI #gry