Search “Reasoning”

Ai2 Releases OlmoEarth v1.3, Offering Global 10-Meter Resolution Satellite Embeddings

The Allen Institute for AI released OlmoEarth v1.3 on October 22, 2026, including the OlmoEarth v1.3 Base model, 128-dimensional and 64-dimensional global precomputed annual embeddings, the login-free browser-based OlmoEarth Embeddings Explorer, and a technical report, all under the OlmoEarth Artifact License. The model is a 128-million-parameter ViT-Base foundation model that processes Sentinel-1, Sentinel-2, and Landsat image time series, outputting a 128-dimensional vector for every 10-meter pixel worldwide. Ai2 states that across 27 evaluations covering six continents, the 128-dimensional embeddings scored highest on 20 of them, while the 64-dimensional version led on 18; it also introduced PLANTEUR, a benchmark for crop types in tropical farmland, where the 128-dimensional embeddings achieved 24.9 mIoU in the benchmark's leave-one-island-out test, higher than Tessera v2's 19.1 and AlphaEarth Foundations' 15.8.

Recursive releases its first model Flint: local coding MoE, 125+ tokens/s on a MacBook

Recursive releases its first model Flint, a MoE with 40 billion total parameters and 2.5 billion activated parameters per token, aimed at local Agent coding, with weights open-sourced under Apache 2.0. With short context, it exceeds 125 tokens per second on an M4 Pro MacBook Pro and over 150 on a 64GB RTX Spark, and supports llama.cpp, Ollama, LM Studio, vLLM, SGLang, and Unsloth. Nearly every stage of its R&D involved the automated research system Eureka Machine, including architecture search, quantizing the output layer to 4 bit with offline distillation, and fusing Metal kernels for M4/M5.

Recursive releases laptop-class coding model Flint, says it was built mainly with AI involvement

AI startup Recursive has released its first model, Flint, a coding model that runs offline on a laptop and is freely downloadable on Hugging Face under the Apache 2.0 license; the company says its development process was mainly carried out with the involvement of its in-house Eureka Machine. Flint is an MoE model that activates only 2.5B of its 40B parameters at a time, generates at about 131 tokens/sec on a MacBook Pro with an M4 Pro, and outperforms other models under 50B parameters on SWE-Bench Verified. Recursive says Flint requires more guidance than large paid coding assistants and recommends breaking large tasks into small steps.

NVIDIA Releases Kumo Relational Relational Foundation Model, Predicts Without Training

NVIDIA has released Kumo Relational, a pretrained relational foundation model that can make predictions directly on enterprise databases in a single forward pass, without task-specific training, feature engineering, or table flattening. Across 41 tasks on 15 databases—RelBench V1/V2, SAP SALT, and 4DBInfer—it outperforms supervised relational deep learning models and gradient-boosted tree pipelines built with expert feature engineering, achieving an 8% higher MRR than the AutoGluon ensemble on SAP SALT. The model is available through NVIDIA's open-source GPU-native structured-data-models library or NVIDIA NIM, and supports databases with 500B+ rows.

Recursive Releases Open-Source Coding Model Flint, Says Most of Development Process Was Completed by AI

Recursive released Flint, its first open-weight coding model, with 40B parameters. CEO Richard Socher said that nearly every stage of the model pipeline was completed by its Eureka Machine system, including a new base training method, the model architecture, and code optimized for specific chips. A WIRED author tested it on a Mac with 48GB of memory, using it to build a command-line to-do app, a web research scraping program, and an RSI educational website, and found its speed and accuracy superior to models of the same size. Recursive has raised $650 million from Google Ventures, Greycroft, Nvidia, and AMD Ventures, at a valuation of $4.65 billion.

Inception Releases Diffusion Language Model Mercury 3, Claiming Speeds Four to Eight Times Faster Than GPT-6 Luna

Inception has released the diffusion language model Mercury 3, claiming it wins 8 of 12 benchmarks and runs four to eight times faster than GPT-6 Luna and Claude Haiku 5.5. In Inception's own testing, Mercury 3 scored 76.84 on IFBench, 15 points higher than GPT-6 Luna; it scored 51.08 on SWE-Bench Atlas, versus 33.33 and 32.26 for its competitors. Claude Haiku 5.5 led on the DeepSWE long-code test and long-text retrieval. The model generates 1,068 tokens per second on Nvidia GPUs, is priced at $0.20 per million input tokens and $0.70 per million output tokens, with a 50% discount during the launch period, and is now available on the Inception platform as well as OpenRouter, Baseten, and Vercel; there are currently no independent tests.

NVIDIA Releases Nemotron 3.5 Series of Open Models, Covering Embedding, Reranking, and Document Parsing

On October 21, 2026, NVIDIA expanded the Nemotron Retriever series, launching Nemotron 3.5 Embed (8B BF16, 1B BF16, 1B NVFP4), Nemotron 3.5 Rerank 8B BF16, and the 631-million-parameter Nemotron 3.5 Parse, aimed at multimodal retrieval and document processing for enterprise agent workflows. NVIDIA says that adding Rerank to the agent retrieval pipeline reduces the token cost per query by about 14% (with the 8B embedding model) and 13% (with the 1B model), respectively, while also improving retrieval accuracy. Parse scores 75.20 on ParseBench and processes 10.5 pages per second on a single H100; the open weights, data, and training recipes have been released on Hugging Face and can be deployed via NVIDIA NIM as well as Baseten, DeepInfra, and FriendliAI.

English translation is being prepared.

Memory Tensor Open-Sources MemT 1.0 Series Memory-Native Models

Memory Tensor has open-sourced and released the MemT 1.0 series of memory-native models, including MemT-1.0-Nano with 36B parameters and 3B activated (based on Qwen3.6-35B-A3B) and MemT-1.0-Flash with 288B parameters and 17B activated (based on DeepSeek-V4-Flash), adding a memory computation pathway parallel to attention in selected Transformer modules. The team built a 34-billion-token memory training corpus, of which 20 billion tokens were used for continued pre-training and 14 billion for post-training. In a comprehensive evaluation across 11 memory benchmarks, Flash and Nano achieved average scores of 50.6 and 47.0, respectively, ranking in the top two among the parametric memory methods evaluated; on 500 SWE-bench Verified tasks, the proportion of tasks that Nano succeeded on all three attempts rose from 54.00% for the no-experience baseline to 62.80%.

MiniMax Releases M3.1 Flash Model, ALE Score of 29.6% Ranks First Among Domestic Models

MiniMax officially released M3.1 Flash, formerly known as Space Bunny (Yutu), which previously topped the charts anonymously on OpenRouter. The model scores 29.6% on Agents' Last Exam, ranking first among domestic models, and scores 43 on the AA leaderboard; it achieves 73.7% on DeepSWE v1.1, 82.2 points on CyberGym, 76.2 points on BankerToolBench, tying for first with Claude Opus 5, and 61.7 points on OSWorld V2. The model adopts a sparse attention architecture, with output speeds of up to 200+ TPS and a median of 170+ TPS, input at 1 yuan per million tokens, output at 4 yuan, and cache reads at 0.1 yuan, with comprehensive usage costs reduced by 70% compared to the previous generation M3.

MiniMax Releases M3.1 Flash, Scoring 29.6% on Agents' Last Exam to Rank First Among Chinese Models

MiniMax has officially released M3.1 Flash, the model previously known as Space Bunny (Yutu) that topped the charts anonymously on OpenRouter. It scores 29.6% on Agents' Last Exam, ranking first among Chinese models. The model achieves 73.7% on DeepSWE v1.1, 82.2 points on CyberGym, and 61.7 points on OSWorld V2, with output speeds of up to 200+ TPS and a P50 maintained at 170+ TPS. Pricing is 1 yuan per million input tokens, 4 yuan for output, and 0.1 yuan for cache reads.

English translation is being prepared.

English translation is being prepared.

English translation is being prepared.

Meta's open-weight model Muse Spark Open launches on Fireworks: 816B parameters, 1M context

Meta's open-weight model Muse Spark Open is available on Fireworks from day one of launch. It is an 816B-parameter MoE model that activates 42B parameters per token, with a 1M-token context window, targeting long-horizon coding and agentic tasks. Fireworks says it ranks first among 75 models on Vals AI's Harvey legal agent benchmark with 25.4%, nearly 4 times that of Claude Opus 5 (6.7%); on the Rogo Big Finance Bench it surpasses Claude Opus 5.5 at every difficulty level. Fireworks validated the weights with its own production stack, scoring 91.1% on AIME26, 89.7% on GPQA Diamond, and 84.3% on SWE-bench Verified, and reproduced consistent results with SGLang.

MiniMax M3.1 Flash released: output 200+TPS, price halved

MiniMax released M3.1 Flash, a lightweight model positioned for Coding/Agent optimization, with a 1 million token context, output speed reaching 200+TPS in everyday scenarios, with P50 stable at 170+TPS. Its input price is 1 yuan per million tokens, output 4 yuan per million tokens, and cache read 0.1 yuan per million tokens, roughly half of DeepSeek V4.1 Flash-1017's peak-hour price. On Automation-Bench v1.0.6 it scored 62.2 points, surpassing DeepSeek's 60.5 points; on DeepSWE v1.1 it scored 73.7 points, lower than DeepSeek's 77.0 points.

MiniMax Claims Blind-Test Model Space Bunny, Confirms It as MiniMax-M3.1-Flash

MiniMax officially tweeted to claim the mysterious model Space Bunny on an overseas blind-test platform, confirming it as the upcoming MiniMax-M3.1-Flash. The model had previously drawn developer attention in blind tests with a speed of about 100 words per second and a one-million-token context window; community tests measured generation speed steady at 85 to 112 words per second, and it can run through complex 3D scenes and code builds in a single pass within three minutes, and respond quickly in multi-file project repairs.

English translation is being prepared.

Google Releases Gemini 4 Argon, Limited-Time Half-Price API

Google has released its flagship model Gemini 4 Argon, focused on coding, professional knowledge work, and sustained execution of complex tasks. The API launch price is $2 per million input tokens and $10 per million output tokens, with the discount running through November 30; from December 1, prices return to $4 and $20. On the third-party Artificial Analysis Intelligence Index, Argon High ties with GPT-6 Astra Max and Claude Fable5.1 Max at 53 points, at an average cost of $1.99 per task, about 39% lower than Astra. Google says that on an ADK project it worked unattended continuously for 11 hours, proposed six optimization plans involving 32 files, and ran 16,892 tests, two of which were merged, reducing long-session memory preparation overhead by about 40%.

English translation is being prepared.