Open Source Models

French AI startup Mistral AI is negotiating a new funding round of around 3 billion euros at a valuation of approximately 20 billion euros.

Google DeepMind's DiffusionGemma is an experimental open-weight model that uses diffusion to generate text blocks in parallel, offering faster local inference compared to traditional autoregressive models. Built on the Gemma 4 26B A4B MoE architecture, it trades some quality for speed, making it ideal for interactive and editing tasks. The article explains its architecture, how text diffusion works, benchmark results, and provides a step-by-step guide to run it locally using llama.cpp.

Cohere has released its first developer-facing coding model, North Mini Code, a 30B total parameter mixture-of-experts model with only 3B active parameters per token. It runs on a single H100 GPU, supports 256K context length, and is optimized for code generation, agentic software engineering, and terminal tasks. The weights are open under Apache 2.0.

Ollama's MLX engine has been updated to deliver its highest performance on Apple Silicon yet. By leaning more heavily on Apple's unified memory and the Metal-backed MLX framework, models output higher quality responses, respond faster, and use less memory. The update includes support for NVFP4 format, up to 20% faster output, and a snapshot system for agent workflows.

Google has released DiffusionGemma, a new open-weight model under Apache 2 license, available for free via NVIDIA's NIM cloud API. It delivers impressive generation speeds exceeding 500 tokens per second.

Google released DiffusionGemma, a 26-billion-parameter model that generates text via diffusion, achieving 1,000 tokens per second on an H100 GPU—four times faster than autoregressive models, but with lower quality. It's currently experimental.

DiffusionGemma is Google DeepMind's experimental open text generation model that uses text diffusion instead of standard autoregressive decoding, achieving up to 4x faster generation on dedicated GPUs. The 26B MoE model (3.8B active parameters) is built on the Gemma 4 backbone, supports multimodal inputs (text, image, video), has a 256K context window, covers 140+ languages, and is released under Apache 2.0.

Google DeepMind released DiffusionGemma, an experimental open model for fast text generation using parallel token generation. NVIDIA optimized it to run faster on GeForce RTX, RTX PRO, and DGX Spark systems, achieving up to 1000 tokens/sec locally.

NeuroBait is a fine-tuned AI model designed to help ADHD brains by providing dopamine sparks to overcome task initiation paralysis. Created from real observation of the author's wife, it uses warm, flowing prose to offer one tiny actionable step instead of overwhelming to-do lists. Built with LoRA on Gemma 3 12B and deployed on Hugging Face, it aims to help anyone feeling stuck, not just those with ADHD.

A study evaluated LLaMA 3.1 for extracting structured data from Dutch brain MRI reports. The model showed high performance on categorical variables like visual rating scores but lower performance on numerical variables. Few-shot prompting improved numerical extraction accuracy significantly.

Doubleword's batch inference offering keeps costs down by keeping throughput high, something which isn't easily done given the architecture of popular Mixture-of-Expert models. While MoE's sparse expert weights make them quick to train, they also mean that at each layer of every forward each request in a batch typically requires different expert weights to be loaded. This makes inference severely memory-bandwidth bound and cuts throughput relative to dense models. However, by reordering inputs so that similar prompts batch together, we can overlap the experts needed and reduce the number of unique experts loaded per forward.

This article details the configuration and memory calibration required to run the Qwen 3.6 35B MoE model at a 450,000 token context window on a single 32GB VRAM GPU (NVIDIA RTX 5090) using llama.cpp with TurboQuant and YaRN scaling. It covers model selection, quantization trade-offs, KV cache quantization, RoPE scaling, multimodal setup, replication guide, VRAM lifecycle management, and performance evaluation.

Harness-1 is a 20B retrieval subagent built on gpt-oss-20b, trained with reinforcement learning inside a stateful search harness. The harness handles bookkeeping—candidate pool, curated set, evidence graph, verification records—while the policy focuses on search, curation, and verification decisions. It achieves 0.730 average curated recall across eight benchmarks, outperforming the next open subagent by 11.4 points and trailing only Opus-4.6. Weights and harness code are public.

Unlike GPT-4o or Qwen3.5-Omni, Audio Interaction doesn't wait for a recording to end: it translates, transcribes, chats, and picks up everyday noises like coughing in a single stream. Code, model weights, and download instructions are available on GitHub under the Apache 2.0 open-source license, with the training data to follow.

A personal project called tinderbox allows users to export Claude.ai conversations, index them locally, and search them from any Claude session via an MCP server. Supports hybrid retrieval, Supabase storage, and Ollama embeddings.

Google DeepMind released Quantization-Aware Training checkpoints for Gemma 4, targeting edge devices and consumer GPUs. This comparison of BF16, Q4_0 QAT, and the new mobile QAT format focuses on memory footprint, quality preservation, and deployment suitability using published data.

Google releases new Gemma 4 checkpoints optimized with Quantization-Aware Training (QAT) to reduce memory usage and enable local deployment on edge devices and consumer GPUs. The models include a custom mobile quantization format that cuts memory footprint to 1GB for the E2B model.

On June 3, 2026, Google introduced Gemma 4 12B Unified, an open-source multimodal model that understands text, images, audio, and video within a single architecture. It combines a 256K context window with a laptop-friendly design for agentic workflows and local deployment. This article covers its architecture, features, benchmarks, and practical guidance for developers.

At matched accuracy, open-weight LLMs differ substantially in the shape of their error severity distribution — a difference invisible to the scalar error rate. The Errorquake-10k benchmark scores each response on a continuous 0-4 severity scale across 8 domains and 5 difficulty tiers, revealing that severity profiles provide information beyond error rate.

Ollama 0.30 is now available with improved performance and GGUF model compatibility through llama.cpp, augmenting MLX on Apple silicon and supporting more models on wider hardware.

NVIDIA has released Nemotron 3 Ultra, a 550B total (55B active) open Mixture-of-Experts hybrid Mamba-Transformer for long-running agents. It pairs a 1M-token context with up to ~6x higher inference throughput than comparable open LLMs at on-par accuracy, and ships with open weights, training data, and recipes under OpenMDW-1.1.

NVIDIA releases Nemotron 3.5 Content Safety, a unified model combining multimodal input, multilingual coverage, custom enterprise policy enforcement, and auditable reasoning for content safety. Built on Google Gemma 3 4B IT and fine-tuned with LoRA, it supports explicit training in 12 languages with zero-shot generalization to ~140 languages. New features include custom policy enforcement via natural language specifications and a THINK mode for auditable step-by-step reasoning. The model achieves ~85% average accuracy across multiple multilingual and multimodal safety benchmarks while maintaining a compact 4B-parameter size and low latency. NVIDIA also releases a safety dataset with multimodal, multilingual safety reasoning traces.

The model, released under the Apache 2.0 license, is another example of how cloud providers are enabling enterprises to run models on local devices for agentic workflows.

Learn how to perform text classification using locally hosted open-source LLMs like Llama 3, Mistral, and Gemma via Ollama and the Scikit-LLM Python library, all without API costs.

This paper introduces Orli (Ordered Regression of Lines), an end-to-end model that unifies text line detection and reading order prediction as a single image-to-sequence task. Trained on 196,691 pages across ten writing systems, Orli marginally exceeds state-of-the-art on cBAD line detection without dataset-specific training, achieves near-perfect coverage and ordering on multiple reading-order benchmarks zero-shot, and adapts to specialized out-of-domain layouts with limited fine-tuning. Code and weights are open-sourced.

POLARIS is a training recipe for small open-weight models that significantly improves long-form creative writing. By combining a frontier LLM judge with human-reference injection in a GRPO framework, the resulting 9B model matches the performance of much larger models and exhibits strong length generalization.

NVIDIA Nemotron 3 Ultra is a 550 billion parameter (55B active) open model designed for long-running agentic workflows, with 1M token context and NVFP4 optimization, leading in agentic benchmarks and cost efficiency.

Google has released the Gemma 4 12B model, a 12-billion-parameter AI model that can run on consumer laptops with 16GB of RAM, filling a gap in the Gemma 4 lineup between mobile and high-performance models.

Google Deepmind's Gemma 4 12B is an open-source model that processes text, images, and audio natively and runs on laptops with just 16 GB of RAM. It nearly matches the twice-as-large 26B model in benchmarks and ships under an Apache 2.0 license for commercial use.

Google DeepMind has released Gemma 4 12B, a 12-billion-parameter dense multimodal model that eliminates traditional encoders, feeding vision and audio directly into the LLM backbone. It runs locally on consumer laptops with 16 GB RAM, under the Apache 2.0 license. The model natively handles text, images, audio, and video, making it the first mid-sized Gemma with native audio input.

Ideogram releases version 4.0 of its text-to-image model as an open-weight model with native 2K resolution, bounding box control, and improved text rendering. On the DesignArena leaderboard, it ranks first among all open models; only closed systems from OpenAI and Google score higher. Commercial use requires a paid license.

This study evaluates Random Forest and four CNNs (ResNet-50, ResNet-101, EfficientNet-B4, ConvNeXt-Large) for transferable satellite-derived bathymetry over 0-20 m depth using Sentinel-2 imagery. Key design choices include preserving spatial continuity (contiguous reef blocks) and a Smooth Weight Function (SWF)-weighted RMSE loss. Intra-regional RMSE ranges from 1.15-1.92 m (as low as 0.26 m for shallow depths), while cross-regional RMSE is 2.46-2.98 m for deep models. On the MagicBathyNet benchmark, the proposed networks achieve 0.19-0.22 m RMSE, outperforming U-Net and a task-specific transformer with fewer parameters. Multi-temporal imagery and median aggregation reduce noise. Optimized architectures and pretrained weights are released for scalable transfer.

This paper introduces Geometry-Aware Tabular Diffusion (GATD), which augments tabular diffusion denoisers with pairwise angles and lengths computed from column value differences as inputs and auxiliary targets. The MLP instantiation achieves state-of-the-art benchmark performance while using 3.5x fewer parameters on average (up to 25x for classification tasks): on ten datasets, it wins 8/10 Shape, 7/10 Trend, and 9/10 downstream utility (F1/RMSE), reducing Shape and Trend error by 27% and 20%. Default loss weights transfer to GNN and Transformer denoisers, improving Shape on 27/30 and Trend on 25/30 architecture-dataset cells. A matched ablation shows supervision (not extra inputs or capacity) drives the gain. This shows explicit relational supervision is a portable inductive bias for tabular diffusion.

The CVE AI Agent is an autonomous vulnerability intelligence engine that continuously ingests, enriches, and triages CVE data, delivering findings to platforms like n8n, Jira, Slack, Splunk, or local file exports. It features a token-efficient architecture using deterministic minimization logic to filter noise, with prompts averaging 1,000 tokens. The agent follows a strict Two-Pass architecture: Pass 1 extracts all measurable data deterministically, and Pass 2 uses an LLM to fill qualitative sections. It supports multiple LLM providers, including Gemini, OpenAI, Claude, Groq, and Ollama, and offers a web dashboard.

Researchers propose MIND (Data Manifold-aware Image Diffusion Model), which explicitly models manifold geometry by integrating discrete patch tokenization into the score function of a continuous diffusion model. On ImageNet 256×256, MIND-B with 130M parameters achieves FID 2.06 under guidance, surpassing LlamaGen-3B's performance.

Proposes SENSE, which uses target model hidden states for semantic retrieval and soft-gated evaluation to improve robustness and efficiency of retrieval-based speculative decoding, achieving up to 4.09 mean acceptance length and 3.26x speedup on LLaMA and Qwen.

Chinese AI company MiniMax has released its new model M3. It's billed as the first open-weight model to combine top-tier coding performance, a one-million-token context window, and native multimodality.

NVIDIA announced Nemotron 3 Ultra, a 550B-parameter open weights model with 55B active parameters, achieving the highest intelligence among US open weights models with a score of 48 on the AI Index, and serving over 300 tokens per second on DeepInfra.

MiniMax releases M3, the first open-weights model integrating coding & agentic frontier, sparse attention for 1M context, and native multimodality. It achieves strong benchmarks like 59% SWE-Bench Pro, with API pricing starting with a 7-day 50% discount. Weights and tech report due in ~10 days.

This study introduces a multi-model paradigm to study synthetic deception via LoRA fine-tuning of five transformer models. Linear probes detect deception with near-perfect AUC in early layers, and logistic regression probes outperform MLP probes, supporting the Linear Representation Hypothesis. Probes generalize across domains with minimal loss. Different models exhibit distinct representational regimes: collapse in Pythia/Llama/Qwen versus high-dimensional preservation in Gemma-2. The results show that robust, domain-invariant deception representations can be rapidly entrenched through modest supervised fine-tuning, with implications for activation-based monitoring.

MAVEN (Modular Agentic Verification and Execution Network) is a lightweight symbolic reasoning scaffold designed to enhance generalization in tool-calling environments through structured decomposition, adaptive tool orchestration, and intermediate verification. On the MAVEN-Bench stress test, MAVEN improves the GPT-OSS-120b base model from 48% to 71% accuracy without additional training, using an open-weight backbone at roughly 1/10 the cost of proprietary baselines.

An AI agent for long-running, multi-step work and coding. Discussion on Product Hunt.

Ahead of the G7 Summit, the Digital and Technology Ministers' Meeting agreed on definitions for open-source AI, open weights AI, and related terms.

Open-weight AI models without guardrails are becoming more accessible, raising safety concerns. A new method called 'abliteration' easily removes restrictions, allowing anyone to use them for harmful activities like generating terrorist content or making weapons. Despite legitimate uses, regulation and safeguards face challenges.

According to Epoch's internal capability metric (ECI), open-weight models take an average of 4 months to catch up with state-of-the-art closed models. ECI is a composite measure covering many benchmarks.

SnapName is a macOS app that automatically renames screenshots using a bundled local AI model (Gemma 4), ensuring privacy by not uploading images.

At Mistral AI's summit, CEO Arthur Mensch warned that Europe has just two years to build sufficient AI infrastructure or risk becoming a 'vassal state' to American AI. The event drew a large crowd, highlighting growing European demand for data sovereignty and open-source models, despite the region still lagging behind the US in investment and scale.

A project demonstrates boosting Qwen3-30B inference speed from 0.09 to 14.03 tok/s on a 2017 MacBook Air by combining a human experimenter, Codex, llama.cpp, a local database, and IBM Quantum sampling. The QPU is used for candidate selection, not for running the model directly.

NVIDIA's X-Token fixes two structural failures in GOLD and improves GSM8k accuracy from 2.56 to 15.54

Personal insights from the Mistral AI Now Summit: Mistral is evolving from a model company to a full AI stack provider with its own compute, models, platforms, and consultancy. The summit emphasized partnerships (ASML, BNP Paribas, Amazon) over new model releases. Specialized small models (Document AI, Voxtral, Robostral) outperform big general ones for specific tasks. Sovereignty and on-prem deployment are key differentiators for European enterprises. An inspiring talk on using AI to decipher ancient papyrus documents showcased AI's potential in humanities.