Leon56

Daily AI News • Frontier Models • Benchmarks • Tech Updates

[Leon56 @stein_leon27048]

@emreofficialy ACTIVE

[Leon56 @stein_leon27048]

@0xPrajwal_ Na then it is Claude Opus 4.9.1

[Leon56 @stein_leon27048]

GPT-5.6 is live today in three tiers.

Sol: $5 / $30 per M tokens (flagship, 750 tok/s on Cerebras in July) Terra: $2.50 / $15 Luna: $1 / $6

Built for coding and agentic work. ChatGPT paid plans start rollout now.

#OpenAI #GPT5

[OpenAI @OpenAI]

Sol, Terra, and Luna, our GPT‑5.6 family of models, are starting to roll out now in ChatGPT, Codex, and the API.

[Leon56 @stein_leon27048]

Meta ships Muse Spark 1.1 with its first paid API.
$1.25 / $4.25 per M tokens, 1M context, coding + agentic focus.

$20 free credits, US devs in public preview now.

A real price swing at OpenAI and Anthropic.

#MetaAI #MuseSpark

[Leon56 @stein_leon27048]

Robbyant (Ant Group) open-sources LingBot-Video — world's first MoE video foundation model for embodied AI.

30B params, ~3B active per inference.
70k hours of embodied training data.

🔗 https://t.co/oyMAZ94Piy

#AI #Robotics

[Leon56 @stein_leon27048]

Ant Group open-sources LingBot-VLA 2.0 A 6B vision-language-action model trained on 60,000 hours of real robot data across 20 robot bodies. One policy, arms, head, mobile base. Apache 2.0.

#AI #Robotics

[ModelScope @ModelScope2022]

LingBot brings a 6B vision-language-action model to ModelScope, built for practical real-world manipulation across different robot bodies. 🤖 License: Apache 2.0

🔗https://t.co/OPHRgpCWKq
📄 https://t.co/ugSygYs8KW

🏆 On tabletop bimanual manipulation, LingBot-VLA 2.0 reaches 66.2 progress / 34.4 success on Agilex Cobot Magic, ahead of LingBot-VLA 1.0, pi0.5, and GR00T N1.7.

🦾 Trained with 50,000 hours of robot trajectories across 20 robot configurations, plus 10,000 hours of egocentric human videos.

⚙️ Supports arms, end-effectors, grippers, dexterous hands, waist, head, and mobile-base signals in one unified action representation.

🧠 Uses sparse MoE action experts for cross-embodiment scaling, plus dual-query distillation from LingBot-Depth and DINO-Video for current geometry and future dynamics.

[Leon56 @stein_leon27048]

Mistral's first robotics model is here: Robostral Navigate, an 8B VLM that steers robots using just one RGB camera. No LiDAR, no depth sensors, hardware-agnostic.

#AI #Robotics

[Mistral AI @MistralAI]

Announcing Robostral Navigate, our first model for embodied navigation: an 8B robotics navigation model that guides robots to autonomously perform tasks specified with natural language. Single RGB camera. State-of-the-art on R2R-CE.

[Leon56 @stein_leon27048]

xAI's Grok 4.5 runs on a 1.5T-param V9 base, Cursor coding data added via supplemental training. 3x the size of v8-small behind Grok 4. Musk: "close to, perhaps exceeding Opus" in internal evals - no public benchmark yet.

https://t.co/LQfYLM52b3

[SpaceXAI @SpaceXAI]

Announcing Grok 4.5, our first model trained specifically for coding and agents. It was trained with Cursor and offers frontier intelligence at leading speeds and cost efficiency.

https://t.co/i8HpU7w64k

[Leon56 @stein_leon27048]

Cognition launcht SWE-1.7 — das neue Modell hinter Devin.

Trainiert auf Kimi K2.7-Basis, läuft mit 1000 tok/s auf Cerebras-Hardware und kostet $1.97 pro Task. Ab heute in Devin verfügbar (Web, Desktop, CLI).

Quelle: https://t.co/wpD3cZxYRt

#AI #Devin

[Leon56 @stein_leon27048]

Meta launches Muse Image, its first image model from Superintelligence Labs.

Free in Meta AI, Instagram Stories and WhatsApp. Sketch edits, photo blending.

Muse Video previewed, coming soon.

Source: https://t.co/gADel70Vwj

#MetaAI #MuseImage

[Leon56 @stein_leon27048]

@RoundtableSpace Minimax m3 with the 20 Dollar subscription with 1.7b tokens (it‘s fantastic)

[Leon56 @stein_leon27048]

NVIDIA drops compressed Nemotron-3-Puzzle-75B-A9B.

75B total / 9B active. Hybrid Mamba+Attention MoE. NVFP4 on Blackwell, FP8 on Hopper. Open weights on HF. ~2× server throughput vs Nemotron-3-Super.

Source: https://t.co/nB1dTjQ7AF

#NVIDIA #Nemotron

[Leon56 @stein_leon27048]

@msjojos Hello

[Leon56 @stein_leon27048]

@Kingsholz I would Like to

[Leon56 @stein_leon27048]

Sber just dropped GigaChat 3.5 Ultra. 432B MoE (28B active), open weights, hybrid MLA + GatedDeltaNet linear attention. ~40% smaller than the 700B GigaChat 3.1 Ultra. Same hybrid pattern Qwen3.6 ships.

https://t.co/VF7IfVjpLb

#AI #OpenSource

[Leon56 @stein_leon27048]

OpenAI unveils Jalapeño — first custom inference chip, co-built with Broadcom in 9 months.

Purpose-built for LLM inference. Early tests claim substantially better perf/watt vs current state-of-the-art. Deploys late 2026.

Source: https://t.co/8QhfUMXiNm

#OpenAI #Jalapeño

[Leon56 @stein_leon27048]

Google open-sources DiffusionGemma 26B-A4B — first open-weight text diffusion LLM from a top lab.

Generates text in parallel blocks, up to 1,100 tokens/sec on H100. Apache 2.0, runs locally.

Source: https://t.co/vFCN4mMuCA

#Google #DiffusionGemma #OpenSource

[Leon56 @stein_leon27048]

Mistral releases OCR 4 — document AI with bounding boxes, block classification, confidence scores.

170 languages, single-container self-hosting, $4 per 1k pages ($2 batch).

Source: https://t.co/VjhbDJwg0u

#Mistral #OCR #DocumentAI

[Leon56 @stein_leon27048]

Tencent releases Hy3 (Hunyuan 3.0) — open-source MoE, 295B total / 21B active, 256K context, Apache 2.0.

API: ¥1/M input, ¥4/M output, ¥0.25 cached.
Live in Yuanbao, WorkBuddy, CodeBuddy, ima. OpenRouter rollout next.

Source: https://t.co/5L5yQCoj9O

#Tencent #Hy3 #OpenSource

[Leon56 @stein_leon27048]

Transformer co-author Llion Jones + Sakana AI + NVIDIA: open-source CUDA kernels + TwELL sparse format.

>20% speedup on batched inference & training at billion-param LLM scale, 99%+ FFN sparsity.

ICML 2026.

https://t.co/KL43Gb03ev

#AI #ML