Updated
Updated
AIvLLM has published a blog post explaining how its integration with TileRT works, alongside a public benchmark configuration in the InferenceX repository. The benchmark script covers GLM-5.3 FP8 on MI355X hardware and is linked on GitHub. The post itself provides the integration details.
AIThe TileRT and AMD teams reached 469 tok/s single-user decode for GLM-5.3 on 8× MI355X using vLLM. The setup disaggregates work, with vLLM handling prefill and TileRT handling latency-critical decode through vLLM's V1 connector interface. SemiAnalysis's AgentX benchmark reports the configuration at 470 TPS on GLM 5.3 (FP8), over 40% faster than GB300 TRTLLM using FP4.
AIHugging Face released SmolDataEnvs, more than 5,000 open-source RL environments aimed at real-world data science tasks. They target the gap between simple educational games and frontier-level benchmarks, especially for improving coding in models under 10B parameters. The environments are designed as a testbed for developing new RL methods such as GRPO or OPSD.
AIGitHub Security Lab describes the Fuzzing Taskflow, an LLM agent pipeline that identifies entrypoints, writes harnesses, runs AFL++, reads coverage reports, and triages crashes for C/C++ repositories. The agent makes decisions while MCP tools handle execution, and state is stored in a SQLite database. The post also warns that the taskflow runs AFL and build commands directly on the host, so it should be used only in disposable environments without elevated privileges.
Why it matters: The post explains how an LLM agent automates fuzzing steps like harness writing, coverage gap chasing, and crash triage, with a runnable workflow and design tradeoffs.
AIGoogle says Gemma 4 can now run locally on-device within the Antigravity SDK. Developers can build fully local or hybrid multi-agent workflows that pair cloud models with Gemma 4 agents for auditing, patching, and testing code. The post emphasizes total data privacy and zero API fees, powered by LiteRT.
AIOdyssey launched Agora-2, a multi-agent world model that the company is making available for public experimentation. The post predicts such models will increasingly power applications in AI training, AI safety, robotics, autonomous vehicles, defense, energy, cybersecurity, and gaming.
AINVIDIA shared a blog post linking to a story about an open protein dataset, tagging Google DeepMind and EMBL-EBI. The post itself provides no further details about the dataset's contents, scale, or availability.
AIModelScope released Qwen-Image-2.1-Fun-Controlnet-Union, a single checkpoint adding eight structural controls, including Canny, Depth, Pose, and Scribble, plus inpainting to Qwen-Image 2.1. Control and inpainting share one branch with 16 injection points across every second Transformer block, keeping the base model frozen and requiring no checkpoint switching. It runs at guidance scale 1.0 with CFG-distilled sampling and prefix KV caching, and is available under the Qwen Research License with base Qwen-Image 2.1 weights required.
AIGoodfire Research introduces Block-Sparse Featurizers (BSF), which decompose model activations into subspaces rather than single directions. Applied to DINOv3 and Stable Diffusion XL, BSFs find interpretable multidimensional features that better explain activations and enable fine-grained steering. The authors report that most concepts they examined have a stable rank of about two to four dimensions.
AILiquid AI released an experimental DSpark draft model for its LFM2.5-VL-3B vision-language model, delivering decoding throughput gains of up to 2.66× on GPUs and 3.13× on edge devices. The drafter adds about 280M parameters, an 8.9% increase in the deployed model's parameter count, and is available on Hugging Face with support in llama.cpp, SGLang, and MLX-VLM.
AIvLLM now supports Gumbel-max watermarking, which embeds a keyed signal into generated text without changing the expected token distribution. Detection requires the secret key and tokenizer, and the signal accumulates over longer outputs. Benchmarks on Qwen3.5-27B with MTP-3 show throughput changes between -1.1% and +2.0% across batch sizes, with no consistent slowdown.
AIGoogle Developers reproduced Ai2's Olmo 3 7B from scratch in MaxText on Google Cloud TPUs, covering both the stage-1 pre-training run and the stage-2 mid-training anneal. The match was checked on held-out C4 loss, an 8-task accuracy suite, multi-domain perplexity, and token-level KL, not just the training loss curve. The post also describes a data-loader bug that made training loss look better than the reference while held-out metrics did not move.
Why it matters: The post documents how a faithful reproduction was verified on held-out metrics, including a data bug that training loss alone would have hidden.
AIBionic now includes a built-in interactive canvas where users can create Excalidraw diagrams that both they and Bionic can view and edit. The canvas supports collaboration on mockups, system designs, and process maps, and users can ask Bionic to implement what is drawn.
AIJulien Chaumond of Hugging Face posted a thank-you message to supporters of arXiv. The quoted arXiv post reports a $17.2 million multiyear investment from Simons Foundation International, Siegel Family Endowment, and XTX Markets.
AIUnsloth says local AI adoption in 2026 has grown far beyond its own predictions. The company also reports surpassing 500M model downloads on Hugging Face, with Qwen3.8-27B GGUF as its most-downloaded model ever. It says it is working on new features to make local AI faster and more accessible in the coming weeks.
AIUnsloth has passed 500 million model downloads on Hugging Face. Its Qwen3.8-27B GGUF is the most-downloaded model in Unsloth's history.
AIThomas Dohmke argues that Git was built to track code, while the next forge must also capture how that code came to be, with humans and agents working together. He shares this view in a conversation with DevInterrupted, where the context behind code lives alongside it.
AISebastian Raschka argues the main appeal of open-source agent harnesses is not cost but the ability to inspect what they do on users' computers. The post frames transparency over local actions, rather than price, as the key advantage.
AITeleOCR, a lightweight 1.2B vision-language model released under Apache 2.0, parses digital PDFs and warped phone photos without a separate dewarping model. It scores 96.87 overall on OmniDocBench v1.6, the highest among listed specialized VLMs, and ranks #1 in the ICDAR 2026 Sci-ImageMiner Challenge. It supports structured parsing of text, tables, formulas, layouts, and reading order, with synchronous or asynchronous vLLM inference.
AIShanghai AI Lab and SJTU's LUMIA Lab released NCP-ArchPreview, an 8.9B open-weight language model under Apache 2.0. The model reportedly reaches OLMo-3-7B's final Stage 1 loss using 51.3% of the tokens from the 5.73T Dolma 3 corpus, a 1.95× convergence gain. Its concept module jointly predicts tokens and concepts, and domain adaptation updates only its 17M parameters while the token backbone stays frozen.
AIAlibaba's Qwen-Image-2.1 is now the top open-source model in both the Image Edit Arena and Text-to-Image Arena, according to Arena. In the Image Edit Arena it scored 1367 points, placing #16 overall and just 3 points behind GPT-Image-1.5-high-fidelity at #15.
AIUnsloth Desktop received a hotfix update adding image editing for Qwen-Image-2.1. The update also fixes diffusers update issues, GGUF loading failures for Qwen-Image, and black artifacts during diffusion on A100 and consumer GPUs. Users should receive a banner prompting them to update.
AIFireworks released ARCv3, a lossless compressor for BF16 weight-update deltas sent from trainers to RL rollout machines. Across 1,000 production RL deltas, ARCv3 produced payloads nearly 50% smaller than ARCv2, averaging about 0.19% of the BF16 weight size versus 0.36%. ARCv3 is available through the Fireworks Training API as fireworks-delta-compression.
AIBoris Cherny used Opus 5.5 to formally verify the Claude Agent SDK with Lean, and short prompts produced 16 PRs fixing bugs and race conditions. He also combines Lean and TLA+ to find issues in data flow, concurrency, and state management, and says Claude is strong in both languages even though he does not know them well.
AIThe MiMo-V2.6 paper thread reports on the newest open model, which also streams its RL run, focusing on data and RL experimental results rather than architecture. The Pro model reportedly rose from 58.41 to 72.57 on DeepSWE after RL, with the top published DeepSWE score cited at 74.
AIComfyUI's update makes the MiniMax H3 video VAE encode up to about 2.2x faster and decode 1.4-2.7x faster, cutting a 1344x768, 129-frame round trip on an RTX 5090 from 24.3 to 12.7 seconds. The gains come from a fused encoder kernel enabled by default, fp16 accumulation support in a custom convolution, and an int8 decoder, and the source says the changes are visually lossless to the eye. Users need ComfyUI v0.36.0 or above, and the int8 VAE file is a drop-in replacement for the standard one.
AIDaniel Han says Qwen-Image-2.1 works in Unsloth Desktop through INT8, FP8, and GGUF builds, with Unsloth also releasing dynamic GGUFs for it. Pinned RAM offloading lets INT8 and FP8 fit under 6–8GB of VRAM while remaining relatively fast. The linked Unsloth post says the 7B model runs on 12GB VRAM and performs on par with Nano Banana 2.0.
AIUnsloth says the 7B Qwen-Image-2.1 text-to-image and editing model can run locally on 12GB VRAM using its GGUF builds. It also states that the model performs on par with Nano Banana 2.0, and that Dynamic FP8 can run on 6GB of VRAM via offloading for higher quality. The image lists int8 at 7.26 GB with mean LPIPS 0.064 and fp8 at 7.12 GB with mean LPIPS 0.112, and says int8 is the default.
Why it matters: The post gives concrete local-run settings, VRAM figures, and GGUF and FP8 options, which helps readers judge whether the model fits their hardware.
AIXiaomi's MiMo-V2.6-Pro ranks first among open-weight models on the Artificial Analysis Intelligence Index with a score of 46. The author attributes its standing mainly to a training data and post-training recipe that increased agent tasks and used an agentic grader for rewards, rather than its plain Grouped Query Attention and Sliding Window Attention design with a 128-token window.
AIHugging Face's transformers library can now load the same GGUF checkpoints used by llama.cpp, with fast local inference on Mac powered by ggml's Metal kernels. The author, Georgi Gerganov, says the work brings those kernels into the transformers ecosystem to increase compatibility and performance.
AIBlack Forest Labs has released flux-3-action-base, an open-weights 7B world action model that takes camera frames, robot state, and a text instruction to output the next action chunk. The release is an adaptation component rather than a complete robot policy, and new embodiments require their own action heads. The source says the weights are paired with shared video VAE and Qwen3-VL-4B-Instruct text encoders and is governed by the FLUX Kommunity License v.1.0.
Why it matters: The source separates the adaptation base from full robot policies and states the shared encoders and new-embodiment requirements, which clarifies what developers must still build for their robots.
AIOpenBMB thanked a developer for testing MiniCPM5-2B as a worker in a multi-agent workflow handling invoice matching, short payments, duplicate references, and disputes through tool calls. The background post says GPT-6 Astra coordinated the MiniCPM5-2B workers, verifying 32 synthetic invoices in 67.8 seconds with 232 executed tool calls. The demo does not move money.
AIMoonshot's Kimi K3 is now available on Amazon Bedrock, with its license requiring a paid agreement for model-hosting businesses and affiliates above $20M in annual revenue. AWS says customer data stays within its cloud, is not shared with Moonshot or used for training, and inference requests have zero data retention. Neither company disclosed financial terms.
AIQwen-Image-2.1 is now ready to run optimized on Intel hardware with Day-0 OpenVINO support, announced by Intel's developer account. The open-weight checkpoint handles both image generation and editing within a single model.
AIStepFun's Step 5 Preview scores 44 on the Artificial Analysis Intelligence Index at about $0.72 per task, matching Kimi K3 (max) at roughly 2.8x lower cost. The source reports strong reasoning results, including 46% on Humanity's Last Exam, but places it behind Qwen3.8 Max and GLM-5.3 (max) on agentic evaluations. Open weights are planned for October 15.
AITencent Hunyuan says it shrank the 770B-parameter Hy4 preview from roughly 1.5TB to 214 GiB while keeping the parameter count unchanged. The quoted Zhihu post by a Tencent Hunyuan quantization team member describes the method: a 1.25-bit sparse ternary encoding, mixed precision across expert layers, and STQ1_0 CUDA kernels in llama.cpp. The author reports nearly unchanged MRCR retrieval and a small decline in math.
Why it matters: The quoted Zhihu post explains how Hy4 preview's weights were quantized and kept usable at inference, a concrete engineering case for compressing large MoE models.
AIvllm-metal ports vLLM's scheduler, paged KV cache, and OpenAI-compatible server to Apple Silicon, with MLX and Metal handling execution. The v0.28.0 release added batched MTP, GGUF and hybrid-model support, and faster prefill on M5, and v0.29.0 is installable through Homebrew.
Why it matters: The post explains how vllm-metal packs requests and pages KV cache on Apple Silicon, with benchmarks showing where concurrent serving gains and tradeoffs appear.
AI💗 Quoted post (Design Arena): MiMo-V2.6-Pro ranks #8 overall (#3 open-weight) on Design Arena with an Elo of 1338, up 54 points and 22 positions from MiMo-V2.5-Pro. It also ranks #4 overall in Website (#2 open-weight) and #6 overall in Agentic Frontend Development (#2 open-weight).
AIXiaomi MiMo announced open-source releases of Pro and Flash, the MiMo-V2.6-Distill-Qwen-9B model, a technical report, over 7K RL task environments, an end-to-end RL framework, and composable mini-harnesses. The attached table shows MiMo-V2.6-Distill-Qwen-9B after SFT and after RL compared with Qwen3.5-9B, with RL scores higher on most listed benchmarks, such as SWE-bench Verified at 66.2 versus 60.0.
Why it matters: The table compares a 9B distilled model against Qwen3.5-9B on coding, cyber, and agent benchmarks, showing how the reinforcement learning stage changes results.
AIXiaomi MiMo has launched MiMo-V2.6 Pro and Flash, two omnimodal models with open model weights, a technical report, RL environments, and training code. The post says Pro performs on par with Claude Opus 5 and GPT-5.6 Sol across most agent benchmarks and scores 46 on the Artificial Analysis Intelligence Index, the highest among open-source models. A benchmark table compares Pro and Flash with MiMo-V2.5 Pro and frontier models across code agent, general agent, cybersecurity, and visual agent tests.
Why it matters: The source pairs open-weight release details with a benchmark table against Claude Opus 5 and GPT-5.6 Sol, letting readers compare Pro and Flash across agent tasks.