AI Model Release Tracker

• Checkpoints for base, instruction-tuned, and quantized variants now available on Hugging Face Hub and Meta's portal
• Free downloads for local and...

H-JEPA stacks JEPA layers across time scales so high-level plans become low-level subgoals, lifting long-horizon visual planning success from 18% to...

Microsoft quietly uploaded AesCode-32B to Hugging Face on 7 October with zero announcement, a 33B vision-language model that turns prompts into...

OpenAI added gpt-rosalind-discovery to its pricing page under life sciences models next to gpt-rosalind-research, both at $5 per million input and $25...

OrcaRouter released OrcaCyber Zero 1.5 as successor to Zero 1.0, post-trained specifically for vulnerability reproduction, exploit development and...

Two new arXiv papers signal a clear shift in VLM research toward predictive spatial reasoning and visual-native skill learning.

• SpaceCast-Bench...

Meta Muse Code stands out by prioritizing deep repository access and data capture instead of competing on code-completion quality alone.

Public benchmarks offer too little overlap to declare a coding winner between GPT-5.4 and Beam.

• Coding: Beam is not ranked on the public lane, so...

Edda v0.2 large delivers immediate developer value as a fully open Whisper-large-v3 fine-tune for Danish speech recognition, directly loadable via the...

• Gemini 3.8 Live shows lower estimated token cost for chat turns and repository review using listed API rates.
• No shared public benchmarks exist...

Cloudflare released Clef-omni, an open-weight decision model extending the Clef series to native text, image, audio, and video processing.

• Model...

Microsoft's Qwen3.5-9B decision model leads a 36-benchmark comparison at 83.5% average accuracy and 85 ms p50 latency, but ships only as a hosted API with no public weights or disclosed hardware.

SuperNav equips a pretrained MLLM with an agent harness that keeps high-level request interpretation, scene understanding, and decision-making inside...

Reflection AI's new Beam model targets coding and tool-use tasks with a focus on lower inference compute.

• Benchmarks: Scores 80.1 on Terminal Bench...

HeyGen has launched its first in-house voice model, HeyGen Voice, claiming the #1 ranking on ArtificialAnlys for natural, expressive performance. Built for authentic results, it is now available inside the HeyGen product and via API.

JPT-0.8B shows just 6 source-displayable benchmark rows out of 667 tracked slots, all from Decision Index 0.3 categories, with large gaps behind...

• Microsoft launches Microsoft-Decision-1 as a specialized model for quick decisions like AI model routing or image interpretation.
• Positioned after...

Evoke matches a 0.6B dense embedding model on top-100 recall to within 0.004 and finds more relevant results in the top 1,000, all while running on CPUs and directly inside PostgreSQL. The open 30M-parameter model is available today.

• Self Forcing detaches historical KV caches during self-rollout to control memory, but this severs backward gradient paths from later chunks.
  -...

Qwen-Image-2.1-Turbo delivers the same 7B visual generator and Qwen3-VL 8B conditioning in an 8-step checkpoint, trading the base model’s 40-step...