Search “Reasoning”

English translation is being prepared.

English translation is being prepared.

English translation is being prepared.

English translation is being prepared.

English translation is being prepared.

English translation is being prepared.

Former Google DeepMind researcher founds London AI company Metis Reasoning, focusing on compute-efficient reasoning

Metis Reasoning, a London AI company co-founded by former Google DeepMind research lead and AlphaGo co-creator Thore Graepel, has officially launched, seeking a compute-efficient path distinct from the frontier models of Anthropic and OpenAI, and is believed to have raised tens of millions of dollars in initial funding. The company has a team of 12 people, including researchers, mathematicians, robotics experts, and technology entrepreneurs. It believes that LLMs have fundamental limitations in reasoning, that results are difficult to verify, and that there is a lack of audit trails; its approach is more verifiable, more robust, and safer, and it says it has already achieved some positive results on frontier reasoning problems at a cost far below that of frontier models. Metis has not yet announced a release date for its first AI model.

English translation is being prepared.

Stanford CS336 Lecture 10: Percy Liang on Inference Optimization

Stanford CS336 Lecture 10 "Inference" is taught by Percy Liang, who uses a single mathematical thread to connect three layers of inference optimization—architecture, algorithms, and systems: at the architecture layer, GQA/MLA/CLA reduce dimensionality, local attention truncates context, and quantization, pruning, and distillation compress the KV cache and parameters; at the algorithm layer, speculative sampling achieves lossless acceleration; at the systems layer, continuous batching, paging, prefix sharing, and copy-on-write are introduced. The lecture notes are in executable form, with formulas derived step by step by sympy and verified with assertions. Percy believes that the KV cache is the structural root of the Transformer's unfriendliness to inference, and that architectures natively designed for inference (SSMs, linear attention, diffusion-style non-autoregressive generation) may offer the greatest room for improvement.

English translation is being prepared.

English translation is being prepared.

English translation is being prepared.

English translation is being prepared.

Elorian Releases Visual Reasoning Model Glimpse

Elorian has released its first visual reasoning model, Glimpse, which is used to compare scene changes at different points in a video and to review the original clip when changes are subtle or rapid. Elorian calls this frame-based reasoning visual thinking. During reasoning, Glimpse can return to the source video, crop regions, extract specified frames, or run code to measure and compare what it sees; this mechanism is called visual scratch space. Co-founder Andrew M. Dai says it achieves frontier performance competitive with Astra on key public visual reasoning benchmarks.

Jeff Dean praised Elorian for releasing the visual reasoning model Glimpse

Andrew M. Dai announced the launch of Glimpse, calling it Elorian's first visual thinking model, which achieves frontier performance comparable to Astra on key public visual reasoning benchmarks and solves the problem of models only looking at a single static image and failing when encountering motion or viewpoint changes. Jeff Dean reposted and praised the work of AndrewDai and the Elorian team.

Elorian Releases Its First Model Glimpse, Using a Visual Draft Space for Image Reasoning

Elorian has released its first model, Glimpse, opening the API as a research preview, with a waitlist for the chat interface. During reasoning, the model can crop image regions, extract video frames, or run code to measure and compare what it sees, which Elorian calls a visual draft space. Glimpse is priced at $0.60 per million input tokens and $3 per million output tokens, and according to Elorian's evaluations it surpasses GPT-6 Astra on scientific chart reasoning and short video description. The company says the model was trained on several hundred GPUs, a scale notably smaller than the frontier systems it benchmarks against, and it initially targets enterprises with large amounts of video, such as robotics, research laboratories, and media and creative firms.

Aleph Alpha Releases Kolibri-1: A 78B-Parameter Open-Weight Model for German-Language Reasoning

Aleph Alpha has released the open-weight model Kolibri-1, with 78B total parameters. It uses a MoE architecture with 384 experts and 6 routed experts, activating only 3.46B parameters during inference, with a 256K Token context that can be extrapolated to about 1 million Tokens. The model focuses on German and English, and its tokenizer segments German words more token-efficiently; for example, the German Bundestag glossary requires only 1137 Tokens, whereas the compared DeepSeek-V4.1-Flash requires 1910 Tokens. Pretraining used 20 trillion Tokens, with data current as of June 18, 2026. The author's hands-on test found its answers correct and its German-language reasoning process evident, and stated that the Merlin-Arthur protocol makes the model hallucinate less in benchmarks, but its actual performance still needs to be verified in real-world use.

English translation is being prepared.

English translation is being prepared.

StepFun Step 5 Preview demonstrates strong inference performance on NVIDIA GB300

StepFun states that Step 5 Preview demonstrates strong inference performance on NVIDIA GB300 and looks forward to developers building on it. NVIDIA AI, in a quoted tweet, calls this StepFun's largest open-weight release, with a 600B MoE model capable of handling agent tasks up to 24 hours long, reaching nearly 8K tokens/sec/GPU on NVIDIA GB300 with NVFP4, and providing an NVFP4 checkpoint as well as deployment and post-training paths for SGLang, vLLM, NeMo Automodel, and NVIDIA NIM.