Embedl

Embedl was featured in AMD's Advancing AI 2026 keynote! During the headline keynote, AMD announced Embedl as a partner in their Open Robotics Ecosystem, on the "Advancing the Open Robotics Ecosystem" slide under AI Models & Software, alongside names like Nota AI, Liquid, and Ultralytics. Seeing our work presented on that stage is a proud moment for our team. The work behind it: we optimized Physical Intelligence's pi0.5 vision-language-action model to run on AMD iGPUs (Ryzen AI Embedded), bringing robot reasoning to under 100 ms on the edge with no loss in accuracy. That result, 92 ms, made it into AMD's own keynote material. A big thank you to the AMD team for the collaboration and for building a genuinely open, partner-driven ecosystem for robotics. Standing alongside the leaders in AI models, robotic technologies, and hardware is a real milestone for us. Bringing large robotics models to efficient edge hardware is exactly what Embedl exists to do, and AMD's platform makes it possible without compromising on accuracy. More on Advancing AI 2026 in the comments 👇

On July 27, Moonshot AI releases Kimi K3: 2.8 trillion parameters, the first open model in the 3 trillion class. Only 1.8% of those parameters activate per token. That is deliberate. Across the industry, expert pools keep growing while the share doing the work keeps shrinking, because sparsity is what makes models this size affordable to run. Two more things stand out to us: It leads every other open model on essentially all public benchmarks, and matches or beats the top proprietary models on several of them. It introduces a technique called Attention Residuals. An AI model works through a long chain of steps, and until now each step only received one blended summary of everything that came before it. Now each step can go back and pick out exactly what it needs. None of the 2,400 models we have traced does this. The pattern is clear: capability gains at the frontier now come from efficiency. That is the discipline we work in every day at Embedl, making models run on real hardware. It is also why we built HF Viewer, which traces the internals of open models straight from their weights. Over 2,400 models and counting. K3 is next. Full breakdown in the first comment, including an alert for the moment K3 is traced.

The model you validate is not the model that ships. Between the notebook and the vehicle sit compilation, quantisation and a target runtime, and each step can change how the model behaves. If your evidence stops at the float model, you have documented something that never runs in the car. As our Peter Kristiansen puts it in his InCabin interview: the hard part of ISO/PAS 8800 is not reading the clauses, it is producing the evidence continuously. Data science, embedded engineering and verification each hold a piece of the story, and nobody owns the whole chain. In the interview: where teams struggle most, why the notebook and the chip are two different worlds, and the four patterns that separate teams making this work from teams dreading their first audit. Find the link to the article in the comments below 👇 See you at InCabin Europe 2026. #InCabin #AutomotiveAI #EdgeAI #FunctionalSafety

This week we're announcing the HF Viewer Article Editor. You can now write your own interactive articles next to the HF Viewer graph, linking the layers and structures mentioned in the text directly to their positions in the graph. Some really cool articles already written by our users: • LFM2.5-Audio: edge-first speech inference, by Anna Piunova at Liquid AI • DeepSeek V4 mHC Explained, by Shakti Wadekar • Borealis: an open recipe for training an Audio LLM, by Alex Nikolich at Keenable.ai Links to the HF Viewer and all three articles in the comments. We're excited to see the community break down complex models into understandable walkthroughs. Want to write your own? Open any model on hfviewer.com and start writing. We look forward to reading it.

We’re excited to share that Embedl will be exhibiting at InCabin Europe 2026 in Barcelona. Taking place 22nd to 24th September 2026 at La Farga de L’Hospitalet, InCabin Europe brings together the automotive interior intelligence community, including OEMs, Tier 1s, and supply chain engineers working on the future of in cabin technology. We’re looking forward to connecting with industry peers, sharing ideas, and being part of the community driving innovation in automotive interiors. If you’re attending InCabin Europe 2026, we’d love to meet you, visit our booth nr 332 or reach out to Peter Kristiansen directly. #InCabinEurope #Automotive #EmbeddedAI #EdgeAI #InCabin

Embedl SAM 3D Body: image in, 3D body mesh out with end-to-end deployment We've released Embedl SAM 3D Body for TensorRT, a deployable version of Meta's facebook/sam-3d-body-dinov3 single-image human-mesh-recovery model, with the INT8-quantized backbone built with Embedl Deploy: https://lnkd.in/ecsQPw67 Meta's SAM 3D Body reconstructs a full 3D body: pose, shape, camera — from one image. The catch: it's carried by an 840M-parameter DINOv3 ViT-H+/16 encoder. We ship both stages of the pipeline as ONNX, running entirely in TensorRT. The performance (NVIDIA L4, CUDA 13, TensorRT 11, batch 1, 512×512): Encoder (INT8+FP16) → 26 ms — 4× faster and 3.3× smaller than FP32 Decoder + mesh head (FP16) → 16 ms Image → 18,439-vertex mesh → 42 ms / ~24 FPS, with the recovered mesh within ~0.5 % (≈10 mm) of the FP32 result. The backbone is quantization-sensitive and requires mixed precision exploration; the decoder is an iterative promptable transformer that re-runs a parametric mesh rig at every layer. Embedl Deploy's hardware-aware, accuracy-preserving workflow handles the quantization and produces a ready-to-run model. This is not just for SAM 3D Body. The same workflow applies to your models and your hardware. Try running Embedl Deploy on your own models for end-to-end AI deployment on the edge, following our tutorial for SAM3 at https://docs.embedl.com #EdgeAI #TensorRT #NVIDIA #3DHumanPose #ComputerVision #EmbeddedAI #MLDeployment #INT8 #FoundationModels #MachineLearning

Yesterday, one of Embedl's early-stage investors, Fairpoint Capital, invited marketing leaders from their portfolio companies to discuss how the new age of AI is changing GTM strategy. Context matters: every company in the room was at a different stage, in a different industry, with a different motion. Not every insight applies universally. But the direction of travel was clear. - AI is changing how buyers find you. Buyers increasingly use AI tools to research vendors before ever reaching out. AI visibility is becoming as important as traditional SEO, and the rules are fundamentally different. - PR, influencers, and community are the new distribution. Research across 25 million+ citations shows 84% of what AI engines cite comes from earned media. Your own website and paid content contribute almost nothing. Getting covered in credible external sources is now how you get found. - Trust is the new differentiator. Sales cycles are getting longer as information overload grows. Buyers educate themselves more before engaging. Trust wins deals before the first conversation. The qualified lead model is shifting. The classic "book a demo" CTA is losing effectiveness. Surfacing intent earlier and matching it to the right next action is replacing it. - Position on the problem, not the product. Lead with what you solve, not what you do. - Shorter cycles, faster learning. Sprints of six weeks or less are replacing long planning cycles. Understanding what drives results matters more than executing a fixed plan. Thank you to Fairpoint Capital, Ulf Lewander, Susanne Rönnqvist Ahmadi, and all the participants for a great workshop and a lot of interesting insights into GTM in the AI age. #GTM #MarketingStrategy #AIMarketing #B2BMarketing #Embedl

This week, our Chief Sales Officer Richard Savage joined the panel "Can Sweden build defence capability at start-up speed?" at Almedalen, Sweden's annual gathering for politics, business, and society in Visby, Gotland. His take: large organisations in the defence sector can absolutely be agile and innovative. The question is not whether they are capable of moving fast. The question is whether they are working with the right partners to get there. Defence teams need AI that works in the field. That means optimised, safe, and deployable on the hardware that actually exists in operational environments. That is exactly what startups like Embedl are built to deliver. Proud to be part of this conversation. #Almedalen #DefenceAI #EdgeAI #Embedl

Embedl is heading to Visby for Almedalsveckan 2026. We look forward to joining important conversations on innovation, entrepreneurship, technology, and Sweden’s future competitiveness. On Wednesday, Richard Savage will participate in the discussion: Can Sweden build defense capabilities at startup speed? TECHARENA Stage, at 14:20. On Thursday, Richard will also serve as a jury member for the Techarena Challenge from 12:00–14:00. Later that day, Devdatt Dubhashi, Co-founder and Chief Scientist at Embedl, will join Ny Teknik's panel discussion: Customers, capital, talent, or regulations – what is really killing growth? 25 June, 12:45–13:30 Södra Kyrkogatan 8, Visby The discussion will explore why so few Swedish startups make the journey from promising ideas to global growth companies. Is it a lack of capital in the most critical phases, challenges in attracting talent, or regulations and structures designed for large corporations rather than fast-moving startups? Participants: • Devdatt Dubhashi, Co-founder and Chief Scientist, Embedl • Julia Norinder, People & Communications, Zenseact • Karl Bäckström (Ph.D.), CTO & Co-founder, Eneryield • Sara Wallin, CEO, Foundation Chalmers University of Technology • Anna Ljungdahl, Industrifonden If you are attending Almedalen, we would be delighted to connect and discuss the future of AI, innovation, and growth. #Almedalen2026 #Almedalsveckan #Techarena #NyTeknik #DeepTech #SwedenTech

Sverige producerar startups i världsklass. Ändå är det förvånansvärt få som tar sig hela vägen från lovande idé till globalt tillväxtbolag. Alltför många fastnar i den kritiska fasen mellan produkt och lönsamhet – det som ofta kallas "Dödens dal". Är det bristen på kapital i de mest riskfyllda faserna? Ett optionssystem som gör det svårt att konkurrera om talang globalt? Eller regelverk och strukturer som i praktiken är anpassade för stora bolag, inte snabbfotade startups? Under panelen diskuterar vi vad som egentligen håller tillbaka svenska startups, hur nya krav inom områden som AI, data och cybersäkerhet påverkar möjligheterna att växa, och vad som behöver förändras för att fler bolag ska kunna ta steget från lovande idé till internationell framgång. Deltagare: • Devdatt Dubhashi, Chief Sales Officer, Embedl • Julia Norinder, People & Communications, Zenseact • Karl Bäckström (Ph.D.), CTO & Co-founder, Eneryield • Sara Wallin, CEO, Stiftelsen Chalmers University of Technology • Anna Ljungdahl, Industrifonden 📅 Torsdag 25 juni kl. 12.45–13.30 📍 Södra Kyrkogatan 8, Visby