AI Model Release Tracker

Modulate has taken the #1 spot on Hugging Face’s Open ASR Leaderboard out of 88 models, combining leading transcription accuracy with streaming...

Google DeepMind's 2026 output positions it as a direct rival to OpenAI and Anthropic across frontier models and scientific AI.

• Gemini Ultra 2...

GLM-5.2's capture of 40% of OpenRouter developer tokens and the surge of "ZCode alternatives" guides highlight how MIT-licensed open weights fail to...

Ant Group open-sourced SingGuard-NSFA, a guardrail framework that intercepts malicious requests and validates responses for autonomous AI agents...

SpaceXAI’s Grok 4.5, released July 8, has surpassed GPT-5.6-SOL and now approaches Claude Opus 4.8 in browser-use agent tasks while ranking fourth on...

LingBot-Depth converts incomplete noisy depth sensor data into high-quality metric 3D measurements, directly improving robotic perception and 3D mapping accuracy.

Flagship Model Releases

• 🔥 GPT-5.6 Sol Tops Design Arena: OpenAI released GPT-5.6 Sol which secured first place overall on Design Arena with an...

GPT-5.6 Sol claimed first place in Design Arena with an Elo score of 1353, outperforming Claude Fable 5 while matching GLM 5.2. The model delivered an...

• Article 1 shows classic direct model comparisons on coding logic and reasoning tasks, with Opus 4.8 scoring highest on maintainability.
• Article 2...

Open-Source Releases

• 🔥 MuScriptor: Kyutai Labs released MuScriptor, the best open model for multi-instrument transcription to MIDI, in...

DeepSeek V4 Flash (High) scores 64/100 vs GPT-OSS 120B's 34/100 across 13 overlapping benchmarks, with a 30-point margin.

• Agentic: Leads...

MuScriptor, the leading open model for multi-instrument transcription, launches via Kyutai Labs in collaboration with MireloAI. Feed it any genre—pop,...

Zhiyuan's Wujie series, featuring the multimodal world model Emu3 and brain-inspired Brainμ, has surpassed 640 million global downloads, signaling major open-source adoption from this leading Chinese AI lab.

GPT-5-mini surprisingly led the source-relevance F1 benchmark at 0.908, beating larger frontier models on 1,248 human-reviewed decisions.

On factual...

• Coding benchmark: Muse Spark 1.1 earns 71 on the LiveCodeBench-based Intelligence Index sub-score, a 12-point gain from the prior model.
• Agentic...

• Most powerful model yet: Mythos Preview is a full tier above Opus 4.7, leading on every published benchmark including 93.9% SWE-bench Verified and...

Gemini 3.5 Pro reportedly beats Claude Fable 5 and GPT-5.6, scoring 82.4% on GPQA Diamond in leaked internal evaluations. The result sharpens competition in the flagship model race.

GPT-5.4 leads GPT-OSS 120B by 51 points overall (85 vs 34) on 20 shared benchmarks.

• Leads in every overlapping eval including Tau2-Telecom (87.1%...

• MDM technique from LingBot-Depth tackles depth sensing on transparent and reflective surfaces, slashing RMSE from 0.132 to 0.062 in tough indoor...

Tencent released HiLS-Attention, a 7B sparse attention model on Hugging Face that learns chunk selection end-to-end for ultra-long context. The open-source drop highlights efficient scaling without dense attention overhead.