Releases·Model Release
Indie developer releases 3.7B multilingual reasoning MoE trained from scratch on a single RTX 4090
A solo developer uploaded a checkpoint of Polanka 3.7B, a multilingual reasoning mixture-of-experts model pretrained, mid-trained, and fine-tuned from scratch on one consumer-grade RTX 4090 over many months. The checkpoint is public on Hugging Face as a research artifact. It shows small-scale from-scratch pretraining remains feasible on consumer hardware.
57 engagement1 sourceTue, Sep 1, 2026, 03:26 PM
AI Overview
- Trained across pre-training, mid-training, and fine-tuning stages on a single RTX 4090 over many months; the developer says token count was not tracked
- Checkpoint piotr-ai/polanka_3.7b_exp_wip_260901 released on Hugging Face as a WIP research artifact
- Covers 13 languages: Polish, English, Chinese, Czech, Slovak, Ukrainian, Russian, Italian, Spanish, French, German, Portuguese, Lithuanian
- Training data upscaled for Polish, English, and Chinese
Related
Releases·Model Release
mii-llm releases Zagreus-0.4B-por, a small open-source Portuguese-English model
Open-source AI lab mii-llm released Zagreus-0.4B-por, a 400-million-parameter bilingual Portuguese-English language model pretrained from scratch. The model is part of the Zagreus family focused on European languages, with training sponsored by Seeweb and Regolo.ai. It is available on Hugging Face.
Open-source AI lab mii-llm released Zagreus-0.4B-por, a 400-million-parameter bilingual Portuguese-English language model pretrained from scratch. The model is part of the Zagreus family focused on European languages, with training sponsored by Seeweb and Regolo.ai. It is available on Hugging Face.
14 engagement1 sourceredditFri, Jul 24, 2026, 02:37 PM
Social·Community
Users seek sub-3B multilingual models; Qwen 3.5 4B and miniPCM5 1B cited
A Reddit user asks for a multilingual instruction-following model under 3B parameters, noting Qwen 3.5 4B is too large and miniPCM5 1B is weak at multilingual understanding. The thread reflects community interest in compact models for language tasks.
A Reddit user asks for a multilingual instruction-following model under 3B parameters, noting Qwen 3.5 4B is too large and miniPCM5 1B is weak at multilingual understanding. The thread reflects community interest in compact models for language tasks.
28 engagement1 sourceredditSun, Aug 2, 2026, 02:18 PM
Releases·Tool Release
Developer open-sources Deno/WebGPU trainer that pretrains tiny models straight to GGUF
A developer released an open-source trainer that pretrains small language models from scratch directly to GGUF format, eliminating Python/PyTorch dependencies. The trainer runs on Deno with WebGPU and was used to create a 94.7M-parameter base model, Minueza-3-95M-Base.
A developer released an open-source trainer that pretrains small language models from scratch directly to GGUF format, eliminating Python/PyTorch dependencies. The trainer runs on Deno with WebGPU and was used to create a 94.7M-parameter base model, Minueza-3-95M-Base.
3 engagement1 sourceredditTue, Aug 18, 2026, 03:04 AM
Releases·Model Releasefeatured
Thinking Machines releases Inkling-Small, a 276B-parameter MoE model with 1M context window
Thinking Machines released Inkling-Small, a Mixture-of-Experts model with 276B total parameters and 12B active parameters, featuring a 1 million token context window. The model is available in NVFP4 and GGUF formats, with GGUF quantizations provided by Unsloth. This release targets efficient inference with large context lengths.
Thinking Machines released Inkling-Small, a Mixture-of-Experts model with 276B total parameters and 12B active parameters, featuring a 1 million token context window. The model is available in NVFP4 and GGUF formats, with GGUF quantizations provided by Unsloth. This release targets efficient inference with large context lengths.
653 engagement1 sourceredditThu, Jul 30, 2026, 06:01 PM
Releases·Model Release
Aurora-80K releases: an 80K-parameter tiny language model
Aurora-80K, a language model with exactly 80,000 parameters, has been released on Hugging Face. It uses a factorized 4,096-token vocabulary to achieve compact representation. Benchmarks show Wikitext-2 BPB of 3.2902, BLiMP 52.31%, and Arc-Easy 26.05%.
Aurora-80K, a language model with exactly 80,000 parameters, has been released on Hugging Face. It uses a factorized 4,096-token vocabulary to achieve compact representation. Benchmarks show Wikitext-2 BPB of 3.2902, BLiMP 52.31%, and Arc-Easy 26.05%.
246 engagement1 sourceredditThu, Aug 20, 2026, 11:45 AM