Releases·Model Release
Bihar indie developer Abhinav Anand launches Arcle V1, an open-weights 5.84B-parameter omni model, after fraud accusations
A 20-year-old self-taught developer from Bihar, India, who months earlier claimed on Reddit he would build a multimodal model alone, says Arcle V1 is now live with open weights. He says the release follows an anonymous smear site, 'The Dossier', that called the project a fraud. Claims are self-reported and independently unverified in the posts.
11 engagement2 sourcesMon, Sep 7, 2026, 04:24 PM
AI Overview
- Arcle V1 is a single-network omni model handling text, images, documents, speech, and audio in one 2,560-dimensional latent space with open weights on Hugging Face
- The model scores 80.0% on ARC-Easy, 77.5% on GSM8K, 74.2% on MATH-500, and 67.0% on HellaSwag, with a 2,097,152-token context window
- Built by 20-year-old Abhinav Anand alone in Bihar with no team, investors, or CS degree, after an anonymous 'Dossier' site called him a fraud
- The model generates 512×512 images and 24kHz speech and supports 18+ languages, with stated strength in Hindi and India knowledge
- Planned Arcle V2 features text-to-video generation, multi-voice expressive speech, cybersecurity checks, and on-device inference, funded by community donations
Related
Products·launched
A ~511M-parameter fine-tuned model for Rust code assistance running locally on low-end hardware
A hobbyist fine-tuned a small (~511M parameter) LLM specifically for Rust coding. It is free and open source and runs locally on weak consumer hardware like an M1 Mac. It targets solo developers who want offline Rust code help without a large or corporate model.
A hobbyist fine-tuned a small (~511M parameter) LLM specifically for Rust coding. It is free and open source and runs locally on weak consumer hardware like an M1 Mac. It targets solo developers who want offline Rust code help without a large or corporate model.
2 engagement1 sourceredditTue, Sep 1, 2026, 06:27 PM
Social·Community
Developer trains 102M-parameter LLM to test Indonesia's AI foundation model prospects
A developer trained a 102M-parameter language model to explore whether Indonesia can build its own AI foundation model, contrasting with reliance on commercial APIs like ChatGPT and Claude. The experiment highlights feasibility and challenges for domestic AI development.
A developer trained a 102M-parameter language model to explore whether Indonesia can build its own AI foundation model, contrasting with reliance on commercial APIs like ChatGPT and Claude. The experiment highlights feasibility and challenges for domestic AI development.
0 engagement1 sourcerssMon, Aug 17, 2026, 01:29 AM
Products·concept
Unicode-based offline multilingual Q&A system
A concept for a small language model that uses Unicode to detect and answer questions across many languages, including math, music, and art symbols, entirely offline. It aims to provide a lightweight retrieval-based Q&A system that works across domains without internet.
A concept for a small language model that uses Unicode to detect and answer questions across many languages, including math, music, and art symbols, entirely offline. It aims to provide a lightweight retrieval-based Q&A system that works across domains without internet.
1 engagement1 sourceredditTue, Aug 18, 2026, 02:25 AM
Products·building
RHEA: Event-reaction neural architecture for efficient LLM training
RHEA is a novel neural network architecture by a solo developer that replaces standard Transformer layers with event-reaction mechanisms, enabling training of a 1-billion-parameter model on an 8GB RTX 5070 laptop. It aims to make large-scale model training accessible to individuals with limited hardware.
RHEA is a novel neural network architecture by a solo developer that replaces standard Transformer layers with event-reaction mechanisms, enabling training of a 1-billion-parameter model on an 8GB RTX 5070 laptop. It aims to make large-scale model training accessible to individuals with limited hardware.
18 engagement1 sourceredditSun, Aug 23, 2026, 11:55 PM
Releases·Model Release
Tencent open-sources 780B-parameter MoE model with 1M-token context window
Tencent released an open-source 780B-parameter mixture-of-experts model. The model uses 256 experts, supports a 1-million-token context window, and weighs 814 GB even after FP8 quantization. The release gives practitioners access to frontier-scale open weights, though the size makes deployment demanding.
Tencent released an open-source 780B-parameter mixture-of-experts model. The model uses 256 experts, supports a 1-million-token context window, and weighs 814 GB even after FP8 quantization. The release gives practitioners access to frontier-scale open weights, though the size makes deployment demanding.
0 engagement1 sourcerssMon, Aug 31, 2026, 09:21 AM