Perplexity has released pplx-embed-v2-late, a pair of ColBERT-style multimodal embedding models in 0.6B and 9B sizes that retrieve text, images and rendered PDF pages in a shared embedding space. Both are available on Hugging Face under the MIT license, while a hosted API endpoint is planned but not yet live.
we dont do a good job of talking about it but @0xblacklight built a dope effectTS agent harness that 1) stores state as a portable append only log 2) runs on durable objects out of the box 3) has cool retro opentui UI with 10 themes
OpenAI published 722 math manuscripts covering 372 result groups in a new GitHub repository, openai/math, all produced by an unreleased internal model. The author describes the results as including a near-Riemann hypothesis claim pushed to 0.875, and notes that 25 Fields Medal winners criticized the company's approach to AI math research.
Why it matters: The piece traces how AI math results moved from benchmarks to open problems, offering context on verification and the mathematicians' pushback.
The vLLM team released a technical report on vLLM-Omni, a unified serving runtime for omni-modality generation spanning multi-stage autoregressive pipelines, iterative diffusion, and stateful sessions. Current LLM servers and diffusion stacks each cover only one of these patterns, pushing deployments to stitch disjoint runtimes together. vLLM-Omni offers a shared control plane in which an orchestrator advances requests across stages, specialized engines handle compute, and a connector carries payloads.
OpenAI announced on X that it is releasing a series of new mathematical results produced by its internal frontier model. The material is reported as 722 manuscripts posted to GitHub, and the excerpt provides no further detail on the specific results.
Developer Brandon Thomas used Claude Opus 5.5 and Rust to build Artcraft, a free open-source suite with Photocraft, Vectorcraft, and other apps that mimic Adobe products. The author tested Photocraft and found its basic editing commands worked where expected, but Free Transform behaved unpredictably. Thomas describes the software as early alpha and invites developers to contribute.
Google's AI Edge Team open-sourced ML Drift under Apache 2.0, a GPU compute engine for on-device AI inference across OpenGL ES, OpenCL, Metal, and WebGPU. It serves as the core GPU acceleration engine within LiteRT and succeeds the legacy TFLite GPU delegate, which will no longer receive new features. The post cites benchmarks showing up to 40% lower frame latency in YouTube Shorts and up to 30% faster on-device performance in Adobe Lightroom and Photoshop.
Why it matters: The post explains how ML Drift unifies GPU shaders across platforms and replaces the TFLite GPU delegate, which matters for developers deploying on-device models.
🍿🍿🍿🫨
Developer Brandon Thomas's Artcraft has launched seven open source apps in Rust that aim to replicate the interfaces and tools of Adobe Photoshop, Illustrator, Premiere, Lightroom, After Effects, InDesign, and Acrobat Pro. Thomas said he used Anthropic's Claude Opus 5.5 to build the clean-room replacements, with WebAssembly versions available for browser use. The apps remain in a "super early alpha" state, and commenters have pointed out many current shortcomings.
4/ Together, TTFT drops nearly 70% at ~100K throughput. Thanks to @deepseek_ai for the model and kernels, @nvidia for the collaboration, and @SemiAnalysis_ for AgentX. Built by @inferact and the vLLM community.
3/ Kernels: we integrated @deepseek_ai's MegaAttention (NVFP4 KV, 45% smaller), Mega-mHC, Mega-Gate and DeepSelect. Plus vLLM fusions: a CuTe-DSL fused WO-A (up to ~6–7% lower ITL), mHC coefficients on a side stream, and sparse MQA logits (14–23× faster/layer at 512K).
2/ SWA bounded replay: rebuilding sliding-window KV exactly after a prefix hit means replaying 40 × 128 tokens. Bounded replay reruns only the last 128. In prefill, layers 21–39 run only on each request's last 128 tokens. With CUDA graphs, prefill compute drops 30–40%.
Sam Altman reposted an OpenAI announcement that GPT-6 and Intelligent UI are rolling out in ChatGPT for everyone. According to the quoted post, Intelligent UI produces fast, interactive answers with visual explanations and on-the-spot tools for tasks.
This story has a top pick“OpenAI rolls out GPT-6 and Intelligent UI to all ChatGPT users”
🙌 Open models for open science!
Claude Code v2.1.293 adds Claude Haiku 5.5 (claude-haiku-5-5), now the default Haiku model on the Anthropic API, with 1M context and pricing of $0.10/$0.50 per Mtok ($0.50/$2.50 for prompts over 100K). The release also adds agentType to the subagentStatusLine payload and isDeferred to $.tool.register, and fixes numerous issues including a memory leak in HTTP MCP connections.
Create interactive tools you can use right in the conversation, like a game or interactive budget planning calculator.
Preference Model is open-sourcing Karotte, the framework it uses to build reinforcement learning environments that resist reward hacking, including defenses like killing stray processes before grading and rejecting grader-crashing files. The framework has been hardened through more than a million evaluation runs and controlled red-teaming. The company focuses on machine learning engineering tasks for leading labs, and a16z says it is partnering with Preference Model and its founders, Jennifer Zhou and Ning Cao.
With just 2.5GB VRAM, you can train your own Decision Model using small models like Laya. Training is simply done through a UI interface. Video tutorial and analysis are in our guide. GitHub repo: https://github.com/unslothai/unsloth
Reflection AI and Mistral each unveiled new open-source models this week, aiming to beat other Western open models, though they trail top Chinese and closed systems on prominent benchmarks. Reflection CEO Misha Laskin says the target is regulated industries and governments that cannot or will not use Chinese models. The outcome depends on whether businesses and agencies accept less advanced models for some tasks in exchange for lower cost and more control.
Choose d1-3B for high-quality text and vision decisions, or explore our experimental d1-omni-600M when footprint matters. Download the weights, fine-tune, and deploy locally. > Blog: https://www.liquid.ai/blog/d1-open > d1-3B: https://huggingface.co/LiquidAI/d1-3b > d1-omni-600M: https://huggingface.co/LiquidAI/d1-omni-600M > Docs: https://docs.liquid.ai/lfm/models/decision-models
Liquid AI released Open d1, two open-weight multimodal models in its d1 decision model family. The d1-3B model supports text and vision, while d1-omni-600M supports text plus image or text plus audio. The source says the models are meant for real-time decision making across data centers, RTX workstations, and Jetson edge devices.
For more details, check out the full article. See our website as well, to read more about these methodological changes. https://epoch.ai/publications/ebr-bench-update
Liquid AI released two open-weight decision models, d1-3B and d1-omni-600M (experimental), built on its Liquid Foundation Models and available on Hugging Face. d1-3B scores 48.57 on the Decision Index 0.2.1, the highest among decision models under 10B parameters, and answers a question in 16 ms on an NVIDIA Jetson AGX Thor and under 50 ms on a Jetson Orin Nano. The models support text and images (d1-3B) or text with image or audio (d1-omni-600M).
Perplexity is open-sourcing pplx-embed-v2-late, multi-vector embedding models for text and images in one shared space, in 9B and 0.6B sizes. The 9B model can index multimodal data, the 0.6B model can run queries on device, and PDF pages can be searched without OCR. The author reports 92.4% on MADQA and 64% on BrowseComp+, with weights available on Hugging Face.
Unsloth released an open-source method to fine-tune LLMs into decision models that run locally, lifting Qwen3.5 0.8B's aggregate accuracy from 20.7% to 74.3% across three decision benchmarks. The team used a Clef head with LoRA (r=64) for one epoch on just 4GB VRAM, with the approach applicable to models such as Qwen3.8 and Gemma 4. A guide and notebooks are available on the Unsloth documentation site and GitHub.
Both models are on Hugging Face: https://huggingface.co/collections/perplexity-ai/pplx-embed-v2
We're releasing pplx-embed-v2-late, two late-interaction embedding models that retrieve text, images, and pages with a shared embedding space for cross-model querying. Both models achieve frontier performance and are publicly available on Hugging Face. https://www.perplexity.ai/hub/blog/multimodal-embeddings-beyond-a-single-vector
Microsoft Research Asia has open-sourced Agent Lightning v1.0, a roughly 3,500-line agentic RL framework that trains the same agent harness used in deployment. In an end-to-end coding agent pipeline, Qwen3.5-9B rose from 41.8% to 56.4% Pass@1 on SWE-bench Verified using about 6,000 training samples. The framework runs agents as standard Kubernetes jobs without paid commercial sandbox services.
Why it matters: The source shows how training with the deployed agent harness avoids rebuilding agents, and reports concrete SWE-bench Verified gains from about 6,000 samples.
We hope Bolmo opens a path to larger models that adapt how they represent information across languages & domains. Because bytes also represent images & audio, the same ideas could eventually extend beyond text. Learn more in our blog: https://allenai.org/blog/bolmo-nature
you’ll notice the interface is prettttyyyyy slick all that screen real estate for your muse to really spread out in big ups to @ivanbruel and the whole muse team!
AI voice products need to know who spoke and what they said. That is harder than transcription or diarization alone, especially when speakers talk over each other. Pulse by @smallest_AI is now #1 on the Diarization + ASR track of the @voicearena_ai's Diarization Bench, with 24.4% DER. DER is an error rate, so lower is better. The next system in the track scores 40.7%.
Mistral has released Mistral Large 4, nicknamed Le Chonk, a 1 trillion-parameter model it says can be used and customized by anyone. It is in preview, with a final version due by the end of the month, and is optimized for coding and cyberdefense as well as manufacturing, finance, and electrical engineering tasks. Mistral claims it is the most capable open-weight model developed outside China and says it was trained from scratch rather than through distillation.
Stacklok, founded by Kubernetes creators Craig McLuckie and Joe Beda, has released Mecatl, an open source cloud-native harness for coding agents on GitHub. Mecatl keeps the agent loop separate from the client, model provider, state store, and execution environment, and moves tool calling, session management, and memory into manageable systems. The article also covers ToolHive, an MCP platform, and an AI Gateway that is not yet open sourced, with a commercial enterprise control plane tying the pieces together.
The Ballon d'Or, legal division: On Harvey's Legal Agent Benchmark (LAB), Mistral Large 4 is the #1 oss model.
Love this hermes gadgets effort by the community!
Orange AI released "podcast-to-article," a free MIT-licensed open-source skill that reorganizes podcast transcripts by topic rather than speaker timeline. The author says it keeps every number and original quote while cutting off-topic tangents, and that early readership and engagement were good.
Mistral released Mistral Large 4 "Le Chonk", a 1T-parameter (49B active) multimodal model, with open weights planned in about three weeks. Google rolled out Nano Banana 2.1 across Gemini, AI Studio, and the Gemini API, and released EmbeddingGemma 2, a 740M-parameter open multimodal embedding model under Apache 2.0. OpenAI launched the Decisions API in beta with gpt-6-luna, returning typed answers 10x faster than the Responses API.
Meta has open-sourced Rebalancer, a C++ library with a Python interface for solving assignment problems under constraints and objectives, released under Apache 2.0. The article reports that Meta has used it for resource allocation for over 9 years and runs about 40 million problems a day, with P99 solve time of 12 seconds on 265k objects and 3.2k bins. The package can be installed with pip install rebalancer, though PyPI still classifies it as Alpha.
Vela 2.0 brings span-level decisions to routing: safety checks, domain classification, PII spans and unsupported claims in one call. Built by vLLM Semantic Router and KR Labs. Four sizes, 0.3B to 9B. Apache-2.0. https://huggingface.co/collections/vllm-sr/vela-20
InferenceX, the leading open-source ML systems performance dashboard, now has a GTA theme, in addition to CSGO and Minecraft. What themes should we add next? (1/2)🧵