FLUX 3 Klein goes open-weight, runs on 8GB GPUs
Black Forest Labs has released the weights for FLUX 3 Image Klein, a 5.6B-parameter model for text-to-image, editing and multi-reference generation that runs on an 8GB RTX card.
Open Source News and Analysis
152 verified stories covering Open Source, product updates and industry developments. Page 1 of 2
Meta releases 816B open-weight Muse Spark Open
Meta has published the weights of Muse Spark Open, an 816B-parameter MoE model aimed at coding and agents, and says two 512GB Mac Studios can run it.
Kunlun Tech open-sources Matrix-Game 4.0, 60 fps on one RTX 5090
Skywork AI, part of Kunlun Tech, has open-sourced Matrix-Game 4.0: a game engine handles world state while a video model renders frames at 480p and 60 fps on a single RTX 5090.
StepFun open-sources Step 5, a 600B model that runs on 8 GPUs
StepFun released the Step 5 Preview weights on Hugging Face as promised. The 600B MoE model runs in FP8 on a single 8-GPU H200 server.
ModelBest open-sources MiniCPM-V 4.7, a 35B model in 3GB on a phone
ModelBest and OpenBMB release MiniCPM-V 4.7, including a 35B-A3B MoE that reportedly runs in about 3GB of memory on an iPhone 17 Pro Max at over 13 tokens/s.
Small Models May Self-Replicate on TV Boxes Within Months
Anthropic's Frontier Red Team says small open-weight models could be months from copying themselves across cheap IoT devices such as $55 Android TV boxes.
QGIS Votes to Ban AI in Code Contributions
QGIS has merged QEP 428, "No AI contributions permitted," replacing its looser human-in-the-loop policy, though enforcement and non-code content remain unclear.
Cloudflare's Clef-omni Decision Model Now Handles Video and Audio
Cloudflare's new Clef-omni scores audio and full video input in a single pass: about 1.5 seconds for a 21-second clip, at $0.15 per million input tokens.
Claude-built Adobe clones now target Microsoft Office
Brandon Thomas's Rust-based ArtCraft series has grown from seven Adobe-style apps to eleven, adding open-source stand-ins for Word, Excel, PowerPoint and AutoCAD.
JetBrains open-sources Mellum2.1, a small coding agent model
JetBrains has open-sourced Mellum2.1, a 12B MoE model with 2.5B active parameters. It scores 47.0 on SWE-bench Verified, up from 2.0 for its predecessor.
Nvidia open model clears IOI, IMO gold-medal lines
Nvidia's Nemotron 3 models topped gold-medal thresholds at IOI 2026 and IMO 2026, with weights, training data, and inference code released publicly.
vLLM Boosts DeepSeek V4.1 Flash Throughput 5.3x
vLLM's new optimizations push DeepSeek-V4.1-Flash throughput up to 5.3x under rate limits, mainly by shrinking the KV cache and skipping memory offload for long contexts.
Google Open-Sources 740M-Parameter Multimodal Embedding Model
Google open-sourced EmbeddingGemma 2, a 740M-parameter multimodal embedding model for on-device text, image, audio and video search, under Apache 2.0.
Mistral Unveils 1-Trillion-Parameter Large 4 Preview
Mistral's 1-trillion-parameter Large 4 preview scores 38 on Artificial Analysis's Intelligence Index, tying GPT-6 Luna, with open weights due by month's end.
Reflection AI debuts 501B open model Beam
Reflection AI unveils Beam, a 501B-parameter open MoE model claiming to rival Chinese models like GLM 5.2 and Kimi K3, with weights due later this month.
Together Links Open Models Into Claude Code, Claims Half the Cost
Together AI's new Together Link routes Claude Code and Codex sessions to open models from Zhipu, Moonshot AI and DeepSeek, claiming over 50% lower cost than Opus 5.5.
Ai2 Open-Sources 8B Model for Research Reports, 3.5x Faster
Ai2 open-sourced AstaBrief 8B, which writes cited research reports 3.5x faster than a Claude-based pipeline, with similar user approval.
H Company Open-Sources Holo4, a 27B-Parameter Computer-Use Model
French AI startup H Company released Holo4 on Hugging Face under Apache 2.0, a computer-use agent model that operates desktops via screenshots, clicks and keystrokes.
Nvidia Unveils Agent Safety Platform, Backed by 100+ Firms
Nvidia's Open Agent Safety Platform pairs an open-source sandbox runtime with a BlueField-4-based monitor that can halt rogue AI agents in milliseconds.
Colibrì Runs GLM-5.2 on 16GB of RAM by Offloading to SSD
Open-source inference engine Colibrì streams MoE model weights from an NVMe SSD, letting a 16GB-RAM machine run Zhipu AI's 744-billion-parameter GLM-5.2.
vLLM Adds Built-In Text Watermarking With Minimal Throughput Cost
vLLM added a Gumbel-max text watermark toggled by one server flag, with near-zero throughput cost — but code and short replies are hard to detect reliably.
AlphaFold Database Expands to Cover 2,800+ Virus Species
EMBL-EBI, Google DeepMind and Nvidia add protein complex predictions for 2,800+ viruses to AlphaFold, with about 30% never documented before.
GitHub Open-Sources AI Fuzzing Pipeline, Defaults to Sonnet 5
GitHub Security Lab has released an LLM-driven fuzzing pipeline for C/C++ projects that defaults to Claude Sonnet 5, running end-to-end from entry-point discovery to vulnerability reports.
Ant Open-Sources Two 6B AI Design Models, One Splits Into 9 Layers
Ant Group open-sourced two 6B MIT-licensed design models — one generates full UI and poster designs, the other splits a flat image into up to 9 editable layers.
Xiaomi Open-Sources MiMo-V2.6, Scores 46 on Intelligence Index
Xiaomi open-sources MiMo-V2.6, scoring 46 on Artificial Analysis's Intelligence Index after a $3.5 million, six-day reinforcement learning run.
Qwen-Image-2.1 Goes Open, but Commercial Use Needs a License
Alibaba open-sources the image model Qwen-Image-2.1 with 7B parameters and native 2K output, but switches the license from Apache 2.0 to a research-only one, so commercial use now needs a separate agreement.
StepFun Launches Step 5 Preview, Claims Cost 1/8 of Opus 5
StepFun's new Step 5 Preview MoE model claims a top-three open-weight benchmark score and a cost one-eighth of Claude Opus 5, with full weights due October 15.
Claude Speeds Up 30 Biomolecular Models by About 4x
Anthropic had Claude optimize inference for 30-plus open-source biomolecular models, cutting runtime by about 4x on average with minimal accuracy loss; code is on GitHub.
Xiaomi Open-Sources ASR With 12.29% Error Rate for 3-Way Mix
Xiaomi's open-source CocktailASR-1 model targets one speaker's voice in overlapping speech, scoring a 12.29% error rate with three people talking at once.
Open-source search agent Iris ships, context beats raw parameters
AllSpark Research's open-source search agents Iris-mini and Iris-pro post record BrowseComp scores, with context management outperforming raw parameter scaling.
OpenRouter launches US routing, lists Chinese open models
OpenRouter's new US-region endpoint decrypts and processes requests only inside the US, and its list of available models includes Chinese open-weight models DeepSeek V4 Pro, Kimi K3 and GLM 5.2.
Tencent Hunyuan Open-Sources AuK Voice Editing Model
Tencent Hunyuan's 1.5B-parameter AuK model generates, edits and cleans audio from plain-language prompts under an MIT license, with a distilled version running 4.5x faster.
Data Drives 3x More Pretraining Progress Than Model Design
A controlled study finds pretraining data gains outpaced model design gains 12x vs 3.7x from 2019-2025, though annualized rates trail industry estimates.
F-Droid Board Splits Over Adopting Debian's Generative AI Policy
F-Droid proposes adopting Debian's new generative AI policy as an interim rule, but board member Neil Brown publicly objects to AI-written code.
Anthropic open-sources full Lean 4 proof of Fermat's Last Theorem
Anthropic released an Apache 2.0, machine-checked Lean 4 proof of Fermat's Last Theorem built on Mathlib, spanning 60,475 modules and verified by three independent kernels.
Hugging Face Trains a Model to Paint Watercolors With 178 Images
Hugging Face detailed how GRPO reinforcement learning and a 178-image reference pool taught a coding model to paint watercolors for a few hundred dollars.
Moonshot AI Seeks 30% Revenue Cut for Kimi K3
Moonshot AI is negotiating a 30% cloud revenue share for open-weight Kimi K3 with Microsoft, Amazon and Google, testing a first-of-its-kind deal that would beat the industry's old 20% ceiling.
Anthropic Open-Sources Reference Blueprint for Commerce Agents
Anthropic open-sourced anthropics/commerce-agents, arguing skills beat sub-agents, with new rules for caching, security, and memory in production.
Hugging Face Open-Sources 207 Browser GPU Kernels
Hugging Face's new @huggingface/kernels loader ships 207 versioned WebGPU kernels on the Hub, posting a 2.57x geometric-mean speedup across 809 benchmark comparisons.
$399 open-source robot duck rings up $2.6M in its first 24 hours
Pollen Robotics' $399 bipedal robot Microduck sold about 6,500 units worth $2.6 million in a day, running on a seven-process Rust stack with reinforcement-learning-trained walking.
OpenClaw 2.0 Merges 16,000 PRs in One Release
OpenClaw's 2.0 release merges over 16,000 pull requests from 933 contributors in one sweep, reworking setup, browser access, and cloud collaboration.
Tencent Opens Hunyuan Hy4, Narrowly Tops GLM and Kimi in Blind Test
Tencent open-sources Hunyuan Hy4 preview, a 770B MoE model with 6.4% activation and 1M-token context, narrowly beating GLM-5.3 and Kimi K3 in blind tests.
Nvidia Agrees to $12.9B Hugging Face Acquisition
Nvidia has reportedly agreed to acquire Hugging Face for $12.9 billion, nearly doubling last year's rejected $7 billion offer, though the deal remains unsigned and could still fall apart.
SGLang speeds up MiniMax-H3 1.95x with zero quality loss
LMSYS benchmarked MiniMax-H3 on 8x H200 GPUs: SGLang's lossless path beats Diffusers by up to 1.95x, while steeper tiers trade SSIM fidelity for up to 6x speed.
Andrew Ng's OpenWorker Adds Built-In Vulnerability Scanning
Andrew Ng's open-source agent OpenWorker adds a security review role, a fixer-can't-verify rule, and four-tier permissions in its new release.
WeChat's Open-Source Multimodal Embedding Model Tops MMEB-v2
Tencent's WeChat team open-sourced WeMM-Embedding, a multimodal embedding model whose 2B version tops the MMEB-v2 leaderboard, beating rivals four times its size.
Zhipu Open-Sources GLM-5.3-Flash, Cuts Active Params to 18B
Zhipu's open-weight GLM-5.3-Flash halves active parameters to 18B, matches Opus 4.8 on Artificial Analysis, and runs on domestic AI chips at a fraction of the price.
Zhipu Claims Ox Alpha, Usage Tops DeepSeek by 2x
Zhipu confirms Ox Alpha is its new GLM model after anonymous OpenRouter testing, with free-preview usage already more than double DeepSeek's.
Qwen3.8-Flash-Next Launches Tonight, Architecture Jumps Ahead to Qwen4
Alibaba's Qwen3.8-Flash-Next launches tonight on the unreleased Qwen4 architecture, a multimodal MoE with under 5% parameter activation aimed at cutting inference cost.
Meta's Homegrown RDMA Protocol Survives 10% Packet Loss
Meta detailed MetaRoCE, an in-house RDMA transport that skips PFC and still works at 10% packet loss; full spec ships at OCP in October.
Hugging Face Taps Banks to Explore Sale at $13B Valuation
Hugging Face has hired banks to explore a sale near $13 billion, but Bloomberg and TechCrunch say no deal or specific buyer has emerged yet.
Time for Open Models to Catch Up Is Halving Each Generation
SemiAnalysis charts three years of open-vs-closed AI gaps and finds the catch-up window keeps halving, with Kimi K2.6 closing on Opus 4.5 in just 4.8 months.
Ant's Ling-3.0-flash: 2.1x Faster Single-Request Decoding on 4 GPUs
SGLang and Ant Ling Infra cut single-request TPOT for Ling-3.0-flash from 3.33ms to 0.78ms on 4 Blackwell GPUs by removing host-side stalls.
Trillion-Parameter Model Restart Cut From 8.8 Minutes to 32 Seconds
Ant Group and SGLang keep quantized weights resident in GPU memory, cutting Ling-2.6-1T restart time from 8.8 minutes to about half a minute.
Anthropic Embeds Mythos 5 in Security Tools, Grants Open Source $35M
Anthropic folds Claude Mythos 5 into partner security products, funds open-source security with $35M in credits, and widens its Cyber Verification Program.
OpenBMB open-sources 8B model that beats 32B Lean 4 rivals
OpenBMB open-sourced MathForm-8B, an 8B model for autoformalizing math into Lean 4, plus a 367,000-sample verified dataset and evaluation code, outperforming several 32B formalization models.
SenseTime Open-Sources U1.5, 8B Model Outputs Native 4K
SenseTime open-sourced SenseNova-U1.5-8B-MoT with native 4K generation and precise editing, but its parameter count is reported three different ways.
OpenAI Open-Sources Codex's Execution Layer, Cisco Already In
OpenAI open-sourced Codex's execution layer via three integration tiers, with Cisco and a tax-prep firm already using it to build agents.
Liquid AI Cuts Decoding Latency 3.18x With a 300M Draft Model
Liquid AI's LFM2.5-DSpark draft models bring speculative decoding to three on-device models, cutting decoding latency up to 3.18x with identical outputs.
18 AI Models Went Autonomous — Top One Closed 82% of the Gap
Prime Intellect ran 153 fully autonomous research trials across 18 frontier models on a nanoGPT speedrun. The best model closed 82% of the human record gap, but not one of them found a genuinely new method.
Liquid AI revives Q4_0 quantization with 97% accuracy via distillation
Liquid AI's QAD training method restores Q4_0 GGUF checkpoints to 96-97% of BF16 accuracy, preserving the format's speed edge on Arm CPUs and older devices.
Zhipu Opens GLM-5.3 API at $4.40 per Million Output Tokens
Zhipu's GLM-5.3 API prices at $1.40/$4.40 per million tokens and scores 60 on Artificial Analysis's Intelligence Index, ranking 8th of 182 models — but it burns tokens fast.
Qualcomm Open-Sources Mojo Compiler a Month After Modular Deal
Qualcomm has open-sourced Mojo's compiler under Apache 2.0 with the LLVM exception, a month after acquiring Modular — though outside code contributions aren't accepted yet.
DeepSeek Pro adds 384K output
DeepSeek's V4 Pro API now points to the 0813 version, pairing a 1M-token context and 384K output ceiling with a clear price warning.
Alibaba opens Qwen Max-class weights
Qwen3.8-2.4T-A95B has moved from a promise to an open-weight ModelScope repository, giving developers a 4.89TB Max-class model card to test.
Ling tiny model targets local agents
InclusionAI open-sourced Ling-3.0-tiny, a 7.9B-parameter MoE model that activates 1.3B parameters per token and targets local agent workloads.
Meta opens a 30B local agent model
Meta's Muse Glimmer is a 30B open-weight agent model built for local workflows, with Apache 2.0 weights, 24GB and 32GB deployment targets, image input, tool use, and DFlash speculative decoding.
BigBang Lets Models Write Tasks
BigBang-V1 turns synthetic-data production into an agentic loop: models propose, solve, verify, and filter frontier tasks.
Google’s Cyclone AI Adds a Day
Google DeepMind open-sourced WeatherNext Cyclones after Nature results showing roughly one extra day of forecast skill for cyclone track, intensity, and wind structure.
ModelBest Lets Agents Speed Up Industrial Software
ForgeStencil is an OpenBMB system that uses two agents to optimize stencil kernels and deploy verified speedups into real scientific and industrial software.
Inkling Small Closes In on 975B Flagship
Thinking Machines released Inkling-Small with open weights, 276B total parameters, 12B active parameters, multimodal input and benchmark scores close to the 975B Inkling.
MiniMax Opens Its H3 Video Model
MiniMax H3 combines 2K, 15-second video, native stereo audio, RMB 0.8-per-second pricing and a coming open-weight release.
Kimi Valuation Hits $35B
Moonshot AI has reportedly closed more than $3.5 billion in F-round financing at a $35 billion post-money valuation after opening Kimi K3 weights.
Tencent Open-Sources an Inference Speedup Stack
Tencent Hunyuan released AngelSpec, an open-source framework for training speculative-decoding drafters and reducing large-model inference cost.
Robots Get 30,000 Hours of Touch
NeoteAI and Fudan released N0 reports, data and code that put tactile sensing at the center of robot learning.
Kimi K3 Falls Behind in Cyber Tests
UK AISI and US CAISI put Moonshot AI's Kimi K3 through cyber benchmarks, finding strong general-model momentum but a clear gap on exploit chains.
Poolside Releases 118B Open Coding Model
Laguna S 2.1 gives enterprises a self-hosted coding-agent base, trading frontier scores for local control and clearer deployment economics.
MiniCPM puts robot models on device
OpenBMB released MiniCPM-Robot with a 1.5B manipulation model, a 0.9B tracking model, and a deployment path focused on local latency.
Hugging Face discloses AI-agent intrusion
Hugging Face says an autonomous AI-agent campaign exploited dataset-processing paths, while its defenders used local LLM analysis to reconstruct more than 17,000 events.
Kimi K3 arrives with 2.8T parameters
Moonshot AI has launched Kimi K3 with 2.8 trillion parameters, a 1M-token context window and published API pricing, while the open-weight proof still depends on weights and model cards.
SingGuard adds open guardrails for agent actions and multimodal policy checks
SingGuard adds open guardrails for agent actions and multimodal policy checks. The piece reviews the verified facts and why the signal matters beyond one announcement.
Claude-powered Bun rewrite sparks Zig backlash
Claude-powered Bun rewrite sparks Zig backlash. A concise localization of the verified facts and the industry signal behind the story.
Zhipu bets on long-horizon agents after listing
Zhipu bets on long-horizon agents after listing. A concise localization of the verified facts and the industry signal behind the story.
LingBot model predicts before it acts
LingBot model predicts before it acts. A concise localization of the verified facts and the industry signal behind the story.
Ollama raises $65 million for local models
Ollama raises $65 million for local models is reframed for global technology readers, with the key numbers, product claims and open questions kept intact.
LingBot opens a robot vision base model
LingBot opens a robot vision base model is reframed for global technology readers, preserving the key numbers, claims and open questions.
Four Chinese models beat Opus at far lower cost
Four Chinese models beat Opus at far lower cost is reframed for global technology readers, preserving the key numbers, claims and open questions.
GitHub Copilot adds open-weight Kimi K2.7 Code
Model choice, Azure hosting and coding-model price pressure.
Palantir and NVIDIA bring Nemotron into air-gapped AI systems
Sovereign AI, open weights and sensitive government workloads.
Together AI lands $800 million as open-model cloud demand surges
Aramco Ventures led the Series C, valuing the Nvidia-backed open-model infrastructure provider at $8.3 billion.
DeepSeek opens DSpark for faster inference
DeepSeek and Peking University open sourced DSpark, an inference framework that raises generation speed by 60% to 85% and can lift single-GPU throughput as much as 6.6 times.
Meituan opens LongCat-2.0 after Owl Alpha mystery
Meituan identified the high-traffic Owl Alpha model as LongCat-2.0, a 1.6 trillion-parameter MoE model released under the MIT license.
OpenAI’s GPT-5.5-Cyber finds a 23-year-old OpenBSD bug
OpenAI’s GPT-5.5-Cyber finds a 23-year-old OpenBSD bug
OpenAI adds human review to AI-found open-source bugs
Patch the Planet combines Codex Security with Trail of Bits and HackerOne so maintainers receive verified vulnerabilities, severity ratings, patches and tests rather than raw AI reports.
Reflection AI signs a $6.3B compute deal with SpaceX
The young open-source AI lab agreed to rent Nvidia GB300 capacity from SpaceX’s Colossus 2 data center, a deal that can run to 2029 and highlights compute as the scarcest asset.
GLM-5.2 closes in on closed coding models
Z.ai’s open-weight GLM-5.2 scored 74.4 on FrontierSWE, near Claude Opus and above GPT-5.5 in some coding benchmarks, while keeping token prices far lower.
Kimi K2.7-Code puts price before leaderboard glory
Moonshot AI released Kimi K2.7-Code with a HighSpeed mode, strong self-reported tool-use scores and a much lower output price than Claude Opus 4.8.
AI subscriptions strain under heavy users
SemiAnalysis estimates that a fully used $200 ChatGPT Pro plan can represent about $14,000 of API value, exposing the weak economics of all-you-can-use AI pricing.
Microsoft Mirage gives video models spatial memory
Mirage stores latent scene features instead of rendered point clouds, making video world-model generation faster and lighter while improving consistency.
Zhipu opens GLM-5.2 with a 1M-token context
Zhipu's GLM-5.2 pairs a 744B-parameter MoE design with a 1M-token window and a coming MIT-style open-weight release for coding agents.