AI Model Release Tracker

Mistral Large 4 (Le Chonk) positions itself as Europe's strongest open-weight contender outside China, but its edge lies more in deployment control...

Reflection AI's 501B-parameter Beam targets coding and agent workloads with an Apache 2.0 release planned later this month.

• Efficiency angle:...

Kandinsky 6.0 Video introduces a family of diffusion models for synchronized text-to-audio-video generation.

• Variants: Lite (3B parameters) and Pro...

Mistral's new model tests Europe's frontier competitiveness, with CEO Arthur Mensch claiming it outperforms Chinese rivals on cybersecurity—yet no specific benchmarks or model comparisons are provided ahead of today's launch.

• Sparse architecture delivers 501B total parameters with only 23B active, yielding 3-4x inference efficiency over GLM 5.2 and leading Western open...

Reflection's new open-weight Beam model claims to match Z.ai's GLM-5.2 on advanced reasoning benchmarks while using 3-4x less inference compute.
-...

Reflection AI’s Beam delivers competitive coding and agentic results with far lower inference costs than larger open models.

• Scores 80.9 on...

Microsoft is targeting low-latency voice interactions by pairing streaming transcription with natural TTS models.

• MAI-Transcribe-2-Streaming...

Claude Fable 5.1 systematically investigated a reported refund-approval bypass where $729 refunds skipped the $500 manager rule.

• It reproduced...

Kolibri, a new open-weight LLM designed for public administration, outperforms competitors from Alibaba and Nvidia, directly challenging US and Chinese model dominance.

HyperBrowseComp introduces 423 human-validated questions across 13 languages, requiring agents to chain evidence from videos, scanned documents,...

Aleph Alpha dropped Kolibri, a sparse MoE model with 78.1B total parameters but only 3.46B active per token, under Apache 2.0 on October 3.

-...

MotorMind shows general VLMs can drive robots zero-shot by linking mid-level actions to deterministic control and async feedback, bypassing...

Agens Volundr 32B keeps KV cache on only 18 of 72 layers, using linear attention with fixed state for the rest. This delivers 23.9 tok/s at 128K context on two 48 GB GPUs and supports a full 262K window on one 48 GB GPU under Apache-2.0.

FrameMorrow selects historical frames for long-horizon video generation by predicting compact prospective tokens that capture future information...

Can a specialized open-weights model perform practical vulnerability research?

• apex-flash-1 (321.3B, MIT) solved 40 of 60 held-out bug tasks...

NVIDIA Cosmos, an open vision language reasoning model, combined with VSS 3.3 now lets developers build and deploy visual AI agents straight from natural-language prompts, cutting video processing costs through adaptive sampling and streaming NIM.

System One decision models are appearing as lightweight control layers for routing, triage and escalation alongside larger LLMs.

• TokenAI released...

Llama 4 Scout and Maverick weights are available for free download and local/self-hosted deployment under the Llama 4 Community License. This directly benefits developers seeking open-weight models they can run entirely on their own infrastructure.

Reflection's open-weight model, backed by Nvidia compute deals, aims to offer Western users a lower-cost path to proprietary AI systems that rival...