Here's a chronological breakdown of some of the most interesting open models released around October 1st - 31st, 2025:

October 1st:

LFM2-Audio-1.5B (Liquid AI): Low-latency, end-to-end audio foundation model.

KaniTTS-370M (NineNineSix): Fast, open-source TTS for real-time applications.

October 2nd:

Granite 4.0 (IBM): Hyper-efficient, hybrid models for enterprise use.

NeuTTS Air (Neuphonic Speech): On-device TTS with instant voice cloning.

October 3rd:

Agent S3 (Simular): Open framework for human-like computer use.

Ming-UniVision-16B-A3B (Ant Group): Unified vision understanding, generation, editing model.

Ovi (TTV/ITV) (Character.AI / Yale): Open-source framework for offline talking avatars.

CoDA-v0-Instruct (Salesforce AI Research): Bidirectional diffusion model for code generation.

October 4th:

Qwen3-VL-30B-A3B-Instruct (Alibaba): Powerful vision-language model for agentic tasks.

DecartXR (Decart AI): Open-source Quest app for realtime video-FX.

October 7th:

LFM2-8B-A1B (Liquid AI): Efficient on-device mixture-of-experts model.

Hunyuan-Vision-1.5-Thinking (Tencent): Multimodal "thinking on images" reasoning model.

Paris (Bagel Network): Decentralized-trained open-weight diffusion model.

StreamDiffusionV2 (UC Berkeley, MIT, et al.): Open-source pipeline for real-time video streaming.

October 8th:

Jamba Reasoning 3B (AI21 Labs): Small hybrid model for on-device reasoning.

Ling-1T / Ring-1T (Ant Group): Trillion-parameter thinking/non-thinking open models.

Mimix (Research): Framework for multi-character video generation.

October 9th:

UserLM-8b (Microsoft): Open-weight model simulating a "user" role.

RND1-Base-0910 (Radical Numerics): Experimental diffusion language model (30B MoE).

October 10th:

KAT-Dev-72B-Exp (Kwaipilot): Open-source experimental model for agentic coding.

October 12th:

DreamOmni2 (ByteDance): Multimodal instruction-based image editing/generation.

October 13th:

StreamingVLM (MIT Han Lab): Real-time understanding for infinite video streams.

October 14th:

Qwen3-VL-4B / 8B (Alibaba): Efficient, open vision-language models for edge.

October 16th:

PaddleOCR-VL (Baidu): Lightweight 109-language document parsing model.

MobileLLM-Pro (Meta): 1B parameter on-device model (128k context).

FlashWorld (Tencent): Fast (5-10 sec) 3D scene generation.

RTFM (Real-Time Frame Model) (WorldLabs): Real-time, interactive 3D world generation.

October 17th:

LLaDA2.0-flash-preview (Ant Group): 100B MoE diffusion model for reasoning/code.

October 20th:

DeepSeek-OCR (DeepseekAI): Open-source model for optical context-compression.

Krea Realtime 14B (Krea AI): 14B open-weight real-time video generation.

October 21st:

Qwen3-VL-2B / 32B (Alibaba): Open, dense VLMs for edge and cloud.

BADAS-Open (Nexar): Ego-centric collision prediction model for ADAS.

October 22nd:

LFM2-VL-3B (Liquid AI): Efficient vision-language model for edge deployment.

HunyuanWorld-1.1 (Tencent): 3D world generation from multi-view/video.

PokeeResearch-7B (Pokee AI): Open 7B deep-research agent (search/synthesis).

olmOCR-2-7B-1025 (Allen Institute for AI): Open-source, single-pass PDF-to-structured-text model.

October 23rd:

LTX 2 (Lightricks): Open-source 4K video engine for consumer GPUs.

LightOnOCR-1B (LightOn): Fast, 1B-parameter open-source OCR VLM.

HoloCine (Research): Model for holistic, multi-shot cinematic narratives.

October 24th:

Tahoe-x1 (Tahoe Therapeutics): 3B open-source single-cell biology model.

P1 (PRIME-RL): Model mastering Physics Olympiads with RL.

October 25th:

LongCat-Video (Meituan): 13.6B open model for long video generation.

Seed 3D 1.0 (ByteDance): Generates simulation-grade 3D assets from images.

October 27th:

Minimax M2 (Minimax): Open-sourced intelligence engine for agentic workflows.

Ming-flash-omni-Preview (Ant Group): 100B MoE omni-modal model for perception.

LLaDA2.0-mini-preview (Ant Group): 16B MoE diffusion model for language.

October 28th:

LFM2-ColBERT-350M (Liquid AI): Multilingual "late interaction" RAG retriever model.

Granite 4.0 Nano (1B / 350M) (IBM): Smallest open models for on-device use.

ViMax (HKUDS): Agentic framework for end-to-end video creation.

Nemotron Nano v2 VL (NVIDIA): 12B open model for multi-image/video understanding.

October 29th:

gpt-oss-safeguard (OpenAI): Open-weight reasoning models for safety classification.

Frames to Video (Morphic): Open-source model for keyframe video interpolation.

Fibo (Bria AI): SOTA open-source model (trained on licensed data).

October 30th:

Emu3.5 (BAAI): Native multimodal model as a world learner.

Kimi-Linear-48B-A3B (Moonshot AI): Long-context model using a linear-attention mechanism.

RWKV-7 G0a3 7.2B (BlinkDL): A multilingual RNN-based large language model.

UI-Ins-32B / 7B (Alibaba): GUI grounding agent.

@Angaisb_
I asked GPT-5 Codex to use GPT-5 and instead it used GPT-5.1
 
It must be coming soon if even Codex knows about its existence, or maybe it's just a typo, but it'd be such a coincidence