The developer unveiled Inkling 2 Small, a Mixture-of-Experts model with 594 billion total parameters and 17 billion active, supporting image inputs and variable effort levels. The model scored 31 on the Artificial Analysis Intelligence Index, a notable jump from the company's prior generation, with gains attributed to more efficient pre-training and a unified reinforcement learning run. It is currently accessible only through the Tinker platform with closed weights, ahead of a planned Apache 2.0 open-weight release next week pending safety testing.
thinkingmachines.ai
· 2026-11-06
A new continual-learning method called the Sleep framework uses two stages—memory consolidation and dreaming—to help language models retain old knowledge while learning new tasks. Consolidation compresses short-term learning traces into model parameters using on-policy distillation and semantic rewards, while dreaming generates synthetic training examples by randomly activating different mixture-of-experts components and filtering them by gradient-based reward signals.
research.google
· 2026-10-28
BottleCap AI has published a fifth ThinkingCap model, this time trained on Google's Gemma-4-26B-A4B-it, cutting the number of 'thinking' tokens the model uses by up to 69.1% and an average of 38.9% across twelve benchmarks. The company says accuracy drops by only 0.15 percentage points, and that looping behavior fell from 3.6% of responses to 0.8%. A PRO version is also offered, claiming up to 87.6% fewer thinking tokens (62% average) for enterprise customers.
bottlecapai.com
· 2026-10-23
DeepSeek released a point update to its open-weight V4.1 Flash model, a 552B-parameter Mixture-of-Experts system with 16B active parameters and a 1-million-token context window that handles both text and images. The company says the new revision needs only a quarter of the HBM and an eighth of the SSD storage for its KV cache compared with the prior version, using 890 bytes per token entry, and its internal benchmarks place the model between GPT-5.6 Sol and Claude Opus 5, beating the latter on some task tests.
tomshardware.com
· 2026-10-23
DeepSeek has shipped V4.1 Flash 1017, dated October 17, 2026, as a snapshot of its V4.1 Flash model available via the deepseek-flash API. The sparse mixture-of-experts model uses a new Causal Encoder-Decoder architecture, activating 8B parameters on input and 16B on output from a 552B-parameter backbone, with native image understanding trained jointly with text from the start.
openrouter.ai
· 2026-10-17
Liquid AI released LFM2.5-24B-A2B, a 24-billion-parameter mixture-of-experts model with only 2 billion active parameters per token, as the largest and final entry in its LFM2.5 family. The model is reasoning-only, generating an explicit chain of thought, and both base and post-trained open-weight versions are now available on Hugging Face and Liquid's Playground.
liquid.ai
· 2026-10-14
StepFun has released Step 5 Preview, a sparse Mixture-of-Experts model with 27B active and 600B total parameters, now listed on OpenRouter. The model targets agentic workflows, supporting a 1-million-token context window for handling large codebases and documents. StepFun highlights its strength in software engineering, professional knowledge work, and finance-related tasks.
openrouter.ai
· 2026-10-08
Mistral has announced the Large 4, a multimodal AI model that combines extensive language and vision capabilities. It includes 49 billion active parameters and a vision encoder with 1.6 billion parameters, built on a Mixture-of-Experts architecture. The model is designed for broad, general-purpose applications.
docs.mistral.ai
· 2026-10-06
Reflection has unveiled Beam, its first open-weight model with 501 billion parameters, designed for tasks like coding, reasoning, and agentic workloads. The model was pretrained on 23.8 trillion tokens and refined through extensive reinforcement learning using over 100 million rollouts on NVIDIA GPUs, aiming to achieve high efficiency and performance. Beam is currently undergoing final testing before a public release of its weights and technical documentation later this month.
reflection.ai
· 2026-10-05
Aleph Alpha, a German AI company, released Kolibri on 3 October 2026 under the Apache 2.0 license, with weights published on Hugging Face. It is a mixture-of-experts model with 78.1 billion total parameters but only 3.46 billion active per token, trained from scratch on 768 NVIDIA B200 GPUs across data centers in Germany and Finland on roughly 24 trillion tokens, over a fifth of them German.
tej.as
· 2026-10-03
Qualcomm introduced two new flagship smartphone chips, the Snapdragon 8 Elite Gen 6 and Extreme Gen 6, at its Snapdragon Summit, emphasizing on-device AI capabilities. The Extreme version can run a 30-billion-parameter mixture-of-experts model locally, includes a new sensing hub for personalized AI agents, and supports 8K60fps video recording. Motorola announced its Signature 27 phone will use the Extreme chip, with availability planned later this year.
techcrunch.com
· 2026-09-22
Researchers behind the dealignai project published a modified version of DeepSeek's V4.1-Flash model with its safety refusal circuitry surgically removed at the weight level, while preserving core capabilities like reasoning, vision, and its 1M-token context. The team says the checkpoint loads like a standard model with no special code needed, and claims HarmBench testing shows a 100% attack success rate for harmful prompts across both low and high reasoning effort settings, compared to the base model's much lower compliance rate.
huggingface.co
· 2026-09-11
GoKawiil