Open source

AIMistral Small 4 is a 119B-parameter MoE model with 6.5B active per token and a 256k context window, combining instruct, reasoning, and Devstral-style coding in one model. It accepts text and image input, lets users set reasoning_effort per request, and is released under Apache 2.0. The model card reports a 40% latency reduction and 3x throughput versus Mistral Small 3 in its tested setups, and its benchmark chart shows reasoning scores on GPQA Diamond, MMLU Pro, AIME-style text tasks, and MMMU-Pro.

Why it matters: The model card names concrete architecture, context, and licensing details, letting readers compare its reasoning toggle and efficiency claims against other open models.

AIMistral AI released Voxtral Mini 4B Realtime 2602, a multilingual realtime speech-transcription model with 13 supported languages under the Apache 2.0 license. The model has a configurable transcription delay from 240ms to 2.4s, and it matches leading offline open-source models at a 480ms delay. The source says it is optimized for on-device deployment and is currently supported only in vLLM.

Why it matters: The source specifies the 480ms delay operating point, 4B size, Apache 2.0 license, and vLLM serving path, which matter for teams weighing realtime transcription deployment.

AIZ.ai has released GLM-4.7-Flash, a 30B-A3B MoE model that it positions as the strongest model in the 30B class. The model reports SWE-bench Verified 59.2 and τ²-Bench 79.5, and supports local deployment through vLLM and SGLang.

Why it matters: The source lists benchmark scores against Qwen3-30B-A3B-Thinking-2507 and GPT-OSS-20B, letting readers compare the 30B-class MoE model directly with its named rivals.

AIBlack Forest Labs released FLUX.2 [klein] 4B, a 4 billion parameter model that unifies text-to-image generation and image editing with multi-reference support. The source says it runs on consumer GPUs such as the RTX 3090 or 4070 with about 13GB VRAM, and its open weights are available under the Apache 2.0 license.

Why it matters: The source specifies a 4 billion parameter model running on about 13GB VRAM under Apache 2.0, which helps readers judge whether local image generation fits their hardware.

AIBlack Forest Labs has published FLUX.2 [klein] 4B Base, a 4 billion parameter text-to-image model that also supports multi-reference editing. The model is undistilled, is released with open weights under Apache 2.0, and is described as fitting in about 13GB VRAM on cards such as the RTX 3090 or 4070, with reference code available in its GitHub repository and support in ComfyUI and Diffusers.

AIBlack Forest Labs has released FLUX.2 [klein] 9B Base, a 9 billion parameter undistilled rectified flow transformer with open weights for text-to-image generation and multi-reference editing. The model is intended for fine-tuning, LoRA training, and research, and fits in about 29GB VRAM on NVIDIA RTX 4090-class GPUs. A reference implementation is available on GitHub, and the model works with ComfyUI and Diffusers.

AIMoonshot AI released Kimi K2.5, an open-source native multimodal agentic model built by continual pretraining on about 15 trillion mixed visual and text tokens. The model card reports a 1T-parameter Mixture-of-Experts architecture with 32B activated parameters and a 256K context length, and it lists benchmark results against GPT-5.2, Claude 4.5 Opus, Gemini 3 Pro, DeepSeek V3.2, and Qwen3-VL-235B-A22B-Thinking. Weights and code are released under a Modified MIT License, with API access on the Moonshot platform.

Why it matters: The model card gives a full benchmark table against GPT-5.2, Claude 4.5 Opus, and Gemini 3 Pro, useful for comparing open multimodal agent models.

AIFunAudioLLM has released Fun-Audio-Chat-8B, a roughly 8B-parameter large audio language model for natural, low-latency voice interaction, under Apache 2.0. It uses Dual-Resolution Speech Representations with a 5Hz frame rate, which the source says reduces GPU hours by nearly 50%, and it supports English and Chinese.

AIMiniMax has released MiniMax-M2.1 model weights on Hugging Face, with API access on the MiniMax Open Platform and the MiniMax Agent product. The company reports gains over M2 on coding and agent benchmarks such as SWE-bench Verified (74.0) and VIBE average (88.6), and says it outperforms Claude Sonnet 4.5 on multilingual scenarios.

Why it matters: The release pairs open weights with a broad benchmark table against Claude and GPT models, letting readers compare coding and agent claims directly.

AIMiniMax released the technical report and pretrained weights for VTP-Large-f16d64, a visual tokenizer that jointly optimizes contrastive, self-supervised, and reconstruction losses. The model scores 78.2 zero-shot accuracy, 85.7 linear probing, and 0.36 rFID, and its generation performance scales with pretraining compute, parameters, and data. Checkpoint weights were listed as "released very soon" in the source.

AIXiaomi released and open-sourced MiMo-V2-Flash, a Mixture-of-Experts model with 309B total and 15B active parameters, under the MIT license. The company reports 73.4% on SWE-Bench Verified, the top score among open-source models, and inference at 150 tokens per second for $0.1 per million input tokens and $0.3 per million output tokens. It supports a hybrid thinking mode and a 256k context window.

Why it matters: The post gives architecture, speculative decoding speedup, and pricing figures, which help readers judge how the efficiency claims are achieved and what they cost.

AIAlibaba's FunAudioLLM released Fun-ASR-MLT-Nano-2512, an 800M-parameter multilingual speech recognition checkpoint on Hugging Face that supports 31 languages, with emphasis on East and Southeast Asian languages. It is trained on hundreds of thousands of hours of speech and is available through the FunASR toolkit. The source's benchmark tables cover the Fun-ASR family rather than this checkpoint, so no checkpoint-specific accuracy figures are reported.

AITongyi Lab has released Fun-ASR-Nano-2512, an end-to-end speech recognition large model trained on tens of millions of hours of real speech, supporting low-latency real-time transcription across 31 languages. The model, which has 800M parameters, targets industry use such as education and finance and claims 93% accuracy in far-field, high-noise conditions. It is available on Hugging Face and works with the FunASR toolkit.

AIThinking Machines has made Tinker generally available, adding support for advanced vision input models and Kimi K2 Thinking. The update also includes a simpler way to sample from models.

AIApple has released SHARP, a model that generates a 3D Gaussian representation of a scene from a single photograph in less than a second on a standard GPU. The output renders in real time as high-resolution photorealistic views of nearby camera positions, with metric absolute scale, and the paper reports reductions of 25–34% in LPIPS and 21–43% in DISTS versus the best prior model.

AIOpenAI has published weights for a sparse model from Gao et al. 2025, used for qualitative results on bracket counting and variable binding, on Hugging Face under the openai/circuit-sparsity repository. The release includes a standalone Hugging Face implementation that loads the converted model and tokenizer with trust_remote_code and runs sample generation. The project is licensed under Apache License 2.0.

AIAlibaba's FunAudioLLM has released Fun-CosyVoice3-0.5B-2512, a 0.5B-parameter LLM-based text-to-speech model on Hugging Face, with an RL variant also published. The model supports zero-shot voice cloning across 9 languages and 18+ Chinese dialects and accents, with streaming output at latency as low as 150ms. On the source's test-en benchmark, it reports a 2.24% WER and 71.8% speaker similarity, and the RL version reports 1.68% WER.

AIAndrej Karpathy built hn-time-capsule, a tool that feeds each December 2015 Hacker News front-page article and its comment thread to GPT-5.1 Thinking for a retrospective analysis. The project, written with Claude Opus 4.5 in about three hours, processes 930 articles at a cost of about $58 and roughly one hour. Results include prescience and wrongness grades for commenters, and the project is hosted on his website with the intermediate data available for download.

AIARC Prize reports that the top Kaggle entry reached 24% on the ARC-AGI-2 private dataset at $0.20 per task, and that all winning solutions and papers are open source. The top verified commercial model, Opus 4.5 (Thinking, 64k), scored 37.6% at $2.20 per task, while a Poetiq refinement on Gemini 3 Pro reached 54% at $30 per task. The author argues that refinement loops are the main driver of 2025 progress, and says ARC-AGI-3 is planned for early 2026.

Why it matters: The post links 2025 competition results to a broader argument about refinement loops, showing how benchmark outcomes are being read as evidence of AI reasoning progress.

AIApple's CLaRa-7B-E2E is a fully end-to-end unified RAG model that jointly optimizes retrieval and generation, with 16x and 128x document compression. It is trained with end-to-end finetuning using differentiable top-k retrieval and a unified language-modeling objective. The model is available on Hugging Face with example end-to-end inference code.

AIApple has published CLaRa-7B-Instruct on Hugging Face, an instruction-tuned unified RAG model with built-in semantic document compression at 16× and 128× ratios. The model answers instruction-following questions directly from compressed document representations, and its paper, GitHub repository, and transformers usage example are referenced in the release.

AIMoonshot AI released Kimi K2 Thinking, an open-source thinking model that interleaves step-by-step reasoning with tool calls across 200 to 300 sequential invocations. The model is a 1T-parameter mixture-of-experts with 32B activated parameters and a 256k context window, and it uses native INT4 quantization for roughly 2x faster generation. The model card reports benchmark results on HLE, BrowseComp, and other tests, and recommends vLLM, SGLang, or KTransformers for deployment.

Why it matters: The model card gives benchmark tables, quantization details, and deployment settings, letting readers compare Kimi K2 Thinking against GPT-5 and other models on specific tasks.

AIMoonshot AI released Kimi Linear, a hybrid linear attention architecture with 48B total and 3B activated parameters and a 1M-token context length, on Hugging Face. The model card reports up to 6.3x faster TPOT than MLA at 1M tokens and up to 75% lower KV cache needs, and says it outperforms full attention on long-context and RL-style benchmarks.

Why it matters: The model card gives concrete long-context speed and memory figures for a hybrid attention design, useful for judging whether linear attention can replace full attention in practice.

AIMoonshot AI has released Kimi-Linear-Base and Kimi-Linear-Instruct, both 48B total and 3B activated parameters with a 1M context length, on Hugging Face. The models use Kimi Delta Attention in a 3:1 hybrid ratio with global MLA, cutting KV cache by up to 75% and boosting decoding throughput by up to 6x at 1M tokens. The KDA kernel is open-sourced in FLA, and the checkpoints were trained on 5.7T tokens.

Why it matters: The model card gives concrete throughput and KV cache figures for a hybrid attention design, which helps readers weigh its long-context tradeoffs against full attention.

AICognition built and open-sourced blockdiff, a file format that creates block-level diffs of VM disks using only filesystem metadata. The company reports its otterlink hypervisor cut snapshot times from 30 to 60 minutes on EC2 to about 5 to 10 seconds for a 128 GB disk with a 5 GB diff, roughly a 200x speedup. The post also covers the sparse file and copy-on-write concepts behind the approach and why OverlayFS, ZFS, and qcow2 were not chosen.

AICognition launched the official DeepWiki Model Context Protocol server, which is free and requires no login or authentication. It gives programmatic access to ask_question, read_wiki_contents, and read_wiki_structure for GitHub repositories indexed on DeepWiki.com. Private repositories require a Devin account with GitHub connected, and open-source maintainers can apply for $500 in Devin credits.

Why it matters: The source names the three tools and the access path, showing how indexed GitHub repositories can be queried programmatically without login.

AICognition is bringing back its Devin Open Source Initiative, offering $500 in Devin ACU credits to open-source GitHub projects with over 100 forks. Projects below that threshold will still be considered. Eligible projects must have an OSI-approved license and be actively maintained, and maintainers can apply through a linked form.

AICognition has launched DeepWiki, a free public version of its Devin Wiki and Devin Search tools that helps developers understand codebases. Users can view docs for any repo by replacing github.com with deepwiki.com in the URL, and more than 50,000 top public GitHub repositories are already indexed. Private repositories require a Devin account.

AICrossmint said Devin became its top contributor to the open-source GOAT SDK during an initial trial, merging 8 pull requests versus 4 for the next contributor. Examples included a DEXScreener plugin built from a documentation URL and a harder Sui blockchain integration that needed three rounds of feedback and about an hour of human involvement. The company said its results depended on proper training, clear task context, and planned validation, not on treating Devin as superhuman.

AICognition is launching the Devin Open Source Initiative, giving selected open source maintainers 500 free ACUs on a Devin Teams plan as part of Devin's general availability launch. The post shows Devin contributing pull requests to projects including Anthropic's MCP Inspector, Dagger, and nanoGPT, with maintainers still reviewing the results. Devin's GitHub integration forwards PR comments and CI checks to help refine changes, though the company warns a human should still verify final quality.

AILiquid AI reports STAR, an evolutionary algorithm that synthesizes tailored neural network architectures from a numerical genome representation. The authors say it produced hundreds of designs that outperform Transformer and hybrid architectures in quality, with smaller caches and parameter counts, and can optimize for latency on target hardware. The full method is described in the arXiv technical report 2411.17800.

Why it matters: The post explains how evolutionary search over a new architecture design space produced designs beating Transformers and hybrids, giving a concrete method for quality versus latency and memory trade-offs.