Breaking
Rays stun Brewers with 2-run rally in 12th inning and force World Series Game 5 with 4-3 winDOJ fails to indict Cassidy Hutchinson, a key Jan. 6 witnessBolsonaro is granted a visa to enter the U.S. after fleeing Brazil following election lossVal Chmerkovskiy returns to ‘DWTS,’ shares heartfelt moment with Carrie Ann InabaIsrael’s election: What happens next?Rays stun Brewers with 2-run rally in 12th inning and force World Series Game 5 with 4-3 winDOJ fails to indict Cassidy Hutchinson, a key Jan. 6 witnessBolsonaro is granted a visa to enter the U.S. after fleeing Brazil following election lossVal Chmerkovskiy returns to ‘DWTS,’ shares heartfelt moment with Carrie Ann InabaIsrael’s election: What happens next?
Advertisement
AI

Meet Google DeepMind’s Genie 4: Real-Time World Model With Native Audio

AI via MarkTechPost
Quick Read

Google DeepMind has released Genie 4 as a research preview for trusted testers. The Genie 4 world model generates explorable worlds from text and image prompts in real time. It also generates the soundscape: ambient noise, music and character dialogue. Users move with keyboard controls and steer events with plain-language prompts. The bigger change is architectural. A Gemini model now runs inside the simulation loop as a planner. That pattern is important for anyone building training environments for agents or robots. TL;DR Runs on: Google-hosted only (Project Genie web app, Gemini API early access). Performance: Photorealistic worlds with native audio and live text steering. Best: Scene consistency for up to 2 minutes, double Genie 3, per TechCrunch’s briefing. Worst: Sessions last a few minutes, and objects drift after repeated prompt edits. What is Genie 4? Genie 4 is a world model from Google DeepMind that simulates interactive environments frame by frame. It is not a video clip generator. Each frame responds to user input as it arrives. DeepMind states that the model is natively multimodal. Audio, music and dialogue are generated together with the visuals. It is DeepMind’s first real-time world model to generate audio natively. It succeeds Genie 3, announced in August 2025. Genie 3 ran at 720p and 24 frames per second, with visual memory reaching back 1 minute. How does Genie 4 use Gemini? Genie 4 borrows from 2 parts of the Gemini family. Gemini Omni anchors the look and behavior of each world. Other Gemini models handle the reasoning that keeps a simulation coherent. The runtime loop works like this: Interpret: When you type a prompt, Gemini adapts it to fit the running scene. Plan: Gemini 3.8 Flash breaks a goal into a sequence of steps. Verify: It reviews incoming frames to confirm each step succeeded. Adapt: If the world shifts or you intervene, it revises the plan. Between user inputs, Gemini also steers the scene so the world keeps progressing. DeepMind team frames this as an early step toward hybrid world models. In that design, a language model handles structure and narrative, while Genie renders pixels and sound. What can you do inside a Genie 4 world? Steer with text: You can change the weather, move the camera or add a new person or object mid-session. Simulate physics and people: DeepMind cites aerodynamics and collisions. It calls human motion the biggest leap, with characters that react to their surroundings. Hear the world: Every world carries ambient audio. Characters can also hold conversations that react to your inputs in real time. Test robot tasks: In one research demo, Gemini 4 Argon split a checkers task into steps. Genie 4 received each step as a text caption and simulated Apptronik’s Apollo humanoid playing the game. How does Genie 4 compare with other world models? DeepMind released no benchmark tables for Genie 4. The comparison below uses only disclosed specs from each project’s own source. window.addEventListener("message",function(e){if(e.data&&e.data.g4xHeight){var f=document.getElementById("g4x-frame");if(f)f.style.height=e.data.g4xHeight+"px";}}); Who can use Genie 4, and when? Genie 4 is live for trusted testers today. According to DeepMind, it is expected to be available soon to Google AI Ultra subscribers in Project Genie. Early access partners of the Gemini API and Google AI Studio come next. All Genie 4 video and audio will carry SynthID watermarks. DeepMind states that it ran internal safety evaluations before the preview. Testers so far span creative industries, journalism, education, gaming and engineering. Reported uses include: Filmmakers running multiple live ‘takes’ for pre-visualization. Photojournalists turning a single archival photo into a grounded simulation. Game developers testing styles and mechanics before writing engine code. Enterprise partners generating zero-to-one simulation scenarios. What are Genie 4’s limitations? DeepMind lists several open problems on the Genie 4 model page: Visual details and object permanence drift in longer sessions, after complex prompt edits, or during navigation. Interaction is limited to a few minutes, not extended sessions. Real-world locations are not simulated with perfect accuracy. Legible text usually appears only when it is in the input description. Real-time audio accuracy and multi-agent interactions remain unsolved. Key Takeaways Genie 4 generates world, ambient audio and dialogue together, in real time. Gemini 3.8 Flash plans, checks frames and replans inside the loop. No parameters, resolution, frame rate or benchmarks were disclosed. Access starts with trusted testers, then AI Ultra and Gemini API partners. Drift and short sessions remain the main practical limits. Check out the Genie 4 announcement, the Genie 4 model page and the Project Genie waitlist. Also, feel free to follow us on Twitter and don’t forget to join our 150k+ML SubReddit and Subscribe to our Newsletter. Wait! are you on telegram? now you can join us on telegram as well. Need to partner with us for promoting your GitHub Repo OR Hugging Face Page OR Product Release OR Webinar etc.? Connect with us The post Meet Google DeepMind’s Genie 4: Real-Time World Model With Native Audio appeared first on MarkTechPost.

Google DeepMind has released Genie 4 as a research preview for trusted testers. The Genie 4 world model generates explorable worlds from text and image prompts in real time. It also generates the soundscape: ambient noise, music and character dialogue. Users move with keyboard controls and steer events with plain-language prompts. The bigger change is architectural. A Gemini model now runs inside the simulation loop as a planner. That pattern is important for anyone building training environments for agents or robots.

TL;DR

  • Runs on: Google-hosted only (Project Genie web app, Gemini API early access).
  • Performance: Photorealistic worlds with native audio and live text steering.
  • Best: Scene consistency for up to 2 minutes, double Genie 3, per TechCrunch’s briefing.
  • Worst: Sessions last a few minutes, and objects drift after repeated prompt edits.

What is Genie 4?

Genie 4 is a world model from Google DeepMind that simulates interactive environments frame by frame. It is not a video clip generator. Each frame responds to user input as it arrives.

DeepMind states that the model is natively multimodal. Audio, music and dialogue are generated together with the visuals. It is DeepMind’s first real-time world model to generate audio natively.

It succeeds Genie 3, announced in August 2025. Genie 3 ran at 720p and 24 frames per second, with visual memory reaching back 1 minute.

How does Genie 4 use Gemini?

Genie 4 borrows from 2 parts of the Gemini family. Gemini Omni anchors the look and behavior of each world. Other Gemini models handle the reasoning that keeps a simulation coherent.

The runtime loop works like this:

  • Interpret: When you type a prompt, Gemini adapts it to fit the running scene.
  • Plan: Gemini 3.8 Flash breaks a goal into a sequence of steps.
  • Verify: It reviews incoming frames to confirm each step succeeded.
  • Adapt: If the world shifts or you intervene, it revises the plan.

Between user inputs, Gemini also steers the scene so the world keeps progressing. DeepMind team frames this as an early step toward hybrid world models. In that design, a language model handles structure and narrative, while Genie renders pixels and sound.

What can you do inside a Genie 4 world?

  • Steer with text: You can change the weather, move the camera or add a new person or object mid-session.
  • Simulate physics and people: DeepMind cites aerodynamics and collisions. It calls human motion the biggest leap, with characters that react to their surroundings.
  • Hear the world: Every world carries ambient audio. Characters can also hold conversations that react to your inputs in real time.
  • Test robot tasks: In one research demo, Gemini 4 Argon split a checkers task into steps. Genie 4 received each step as a text caption and simulated Apptronik’s Apollo humanoid playing the game.

How does Genie 4 compare with other world models?

DeepMind released no benchmark tables for Genie 4. The comparison below uses only disclosed specs from each project’s own source.

window.addEventListener(“message”,function(e){if(e.data&&e.data.g4xHeight){var f=document.getElementById(“g4x-frame”);if(f)f.style.height=e.data.g4xHeight+”px”;}});

Who can use Genie 4, and when?

Genie 4 is live for trusted testers today. According to DeepMind, it is expected to be available soon to Google AI Ultra subscribers in Project Genie. Early access partners of the Gemini API and Google AI Studio come next.

All Genie 4 video and audio will carry SynthID watermarks. DeepMind states that it ran internal safety evaluations before the preview.

Testers so far span creative industries, journalism, education, gaming and engineering.

Reported uses include:

  • Filmmakers running multiple live ‘takes’ for pre-visualization.
  • Photojournalists turning a single archival photo into a grounded simulation.
  • Game developers testing styles and mechanics before writing engine code.
  • Enterprise partners generating zero-to-one simulation scenarios.

What are Genie 4’s limitations?

DeepMind lists several open problems on the Genie 4 model page:

  • Visual details and object permanence drift in longer sessions, after complex prompt edits, or during navigation.
  • Interaction is limited to a few minutes, not extended sessions.
  • Real-world locations are not simulated with perfect accuracy.
  • Legible text usually appears only when it is in the input description.
  • Real-time audio accuracy and multi-agent interactions remain unsolved.

Key Takeaways

  • Genie 4 generates world, ambient audio and dialogue together, in real time.
  • Gemini 3.8 Flash plans, checks frames and replans inside the loop.
  • No parameters, resolution, frame rate or benchmarks were disclosed.
  • Access starts with trusted testers, then AI Ultra and Gemini API partners.
  • Drift and short sessions remain the main practical limits.

Check out the Genie 4 announcement, the Genie 4 model page and the Project Genie waitlist. Also, feel free to follow us on Twitter and don’t forget to join our 150k+ML SubReddit and Subscribe to our Newsletter. Wait! are you on telegram? now you can join us on telegram as well.

Need to partner with us for promoting your GitHub Repo OR Hugging Face Page OR Product Release OR Webinar etc.? Connect with us

The post Meet Google DeepMind’s Genie 4: Real-Time World Model With Native Audio appeared first on MarkTechPost.

Source: MarkTechPost

Advertisement

Related stories

AI

Polymathic AI and Princeton Researchers Release FlashMap, a Flow Map Drafter for 6.14x Faster LLM Decoding

Researchers from Polymathic AI, Princeton University, USC ISI, NYU and the Flatiron Institute have released FlashMap. FlashMap speculative decoding replaces the discrete diffusion drafter with a continuous flow map. It drafts a full 16-token block in 1 forward pass. Tokens in that block stay consistent with each other. On Qwen3-8B, the research team reports up to 6.14x lossless speedup on GSM8K. Drafter weights for Qwen3-8B and Qwen3-4B are on Hugging Face, with vLLM and SGLang plugins. TL;DR Size: 0.5B-parameter drafter (500M for Qwen3-8B, 475M for Qwen3-4B); drafts blocks of K = 16 tokens. Runs on: the same GPU as the target model. Paper speedups use 1 NVIDIA H200 NVL at concurrency 1, in bf16. Performance: fastest drafter tested on Qwen3-8B, about 7% above a retrained DFlash on average. Best: 6.14x speedup on GSM8K at temperature 0 (acceptance length 7.69 tokens). Worst: 2.28x speedup on Alpaca at temperature 1. On Qwen3-4B (temperature 0), it trails DFlash on 4 of 9 tasks. What is FlashMap? FlashMap is a speculative decoding drafter that maps Gaussian noise to a token block in 1 step. A small drafter proposes K tokens. The target model then checks all K in 1 parallel pass and keeps the longest matching prefix. Output stays exactly what the target would produce. Speed depends on 2 things. How many drafted tokens get accepted, and how cheaply they were drafted. EAGLE-3 drafts autoregressively, so cost grows with block length. DFlash drafts the full block in 1 pass with block diffusion. Why do diffusion drafters lose acceptance? DFlash predicts each slot in the block independently. The research team calls this the factorized approximation. Picture a continuation that is “new york” or “san diego” with equal odds. An independent drafter proposes “new diego” or “san york” half the time. The target rejects those at slot 2. FlashMap builds on Flow Map Language Models. Each noise sample maps to 1 full block. The noise picks the mode for every slot jointly. How does FlashMap work? The drafter is a 6-block diffusion transformer (DiT), with embedding dimension 2048 and 16 attention heads. It reads hidden states from target layers 1, 9, 17, 25 and 33. Embedding matrix and LM head are copied from the target and stay frozen. Training runs in 2 stages: Flow teacher: a continuous flow over one-hot token blocks, trained with cross-entropy. A time reparameterization handles Qwen3’s 151,936-token vocabulary. Flow map student: semigroup distillation turns the flow into a 2-time denoiser. It jumps from noise to data in 1 step. Both drafters train on 500K samples from Open-PerfectBlend. Total compute across all models and ablations was about 4,600 GPU hours on H200s. window.addEventListener("message",function(e){if(e.data&&e.data.fmxHeight){var f=document.getElementById("fmx-frame");if(f&&e.source===f.contentWindow)f.style.height=e.data.fmxHeight+"px";}}); How does FlashMap compare on benchmarks? The research team retrained EAGLE-3 and DFlash on the same data and settings. Tests cover 9 benchmarks across math, code and chat, at temperatures 0 and 1. Speed: 1-step FlashMap is the fastest drafter on every Qwen3-8B task, at both temperatures. Accuracy: 1-step FlashMap beats DFlash on acceptance length in 15 of 18 settings. It has about half the parameters. Chat gains are largest: at temperature 0, gains over DFlash are 5.5% on math, 11.1% on chat. 2-step mode: acceptance length beats DFlash in 18 of 18 settings, by 10% on average. EAGLE-3: at temperature 1, it has 9% higher acceptance length but runs 35% slower. A shuffle test checks whether drafts use real inter-token structure. At 2 steps, FlashMap recovers most of the joint-structure gain of a 16-step teacher. At 1 step, it keeps only a small fraction. Drafting 4 noise samples as a tree raised speedup by 11% on average. Comparison: FlashMap vs other speculative decoding drafters Average speedups are computed by Marktechpost from Figure 1(c) of the FlashMap paper. Matched rows use the authors’ retrained baselines. How do you deploy FlashMap? The flashmap package needs only torch and transformers>=4.56. Install it with uv pip install flashmap. A spec_generate call runs drafter and target together on 1 GPU. For serving, the project page gives 2 paths: vLLM 0.31: pass the drafter in --speculative-config with num_speculative_tokens set to 16. SGLang 0.5.20: use --speculative-algorithm DSPARK with the FlashMap draft path. Set num_steps=2 to trade some speed for higher acceptance. For sampling above temperature 0, vLLM needs "draft_sample_method": "probabilistic". Key Takeaways FlashMap drafts 16 tokens jointly in 1 pass with a continuous flow map. Up to 6.14x lossless speedup on Qwen3-8B, using a 0.5B drafter. Faster than a retrained DFlash on all 9 Qwen3-8B benchmarks. 2-step mode keeps acceptance flat across the block. Weights are Apache-2.0, code is MIT, with vLLM and SGLang plugins. Check out the paper, project page, GitHub repo and model weights. Also, feel free to follow us on Twitter and don’t forget to join our 150k+ML SubReddit and Subscribe to our Newsletter. Wait! are you on telegram? now you can join us on telegram as well. Need to partner with us for promoting your GitHub Repo OR Hugging Face Page OR Product Release OR Webinar etc.? Connect with us The post Polymathic AI and Princeton Researchers Release FlashMap, a Flow Map Drafter for 6.14x Faster LLM Decoding appeared first on MarkTechPost.

via MarkTechPost
AI

Meet LEANN: An Open-Source Vector Index that Shrinks a 60M-Chunk RAG Index from 201 GB to 6 GB

UC Berkeley’s Sky Computing Lab built LEANN, an open-source vector index that stores almost no embeddings. The LEANN vector index recomputes embeddings at query time instead of saving them to disk. The repo benchmark shows a 60 million chunk Wikipedia index dropping from 201 GB to 6 GB. The research paper won the Best Research Paper award at MLSys 2026. So, it brings semantic search and RAG over personal data, on a laptop. TL;DR Size: 6 GB for 60M chunks (repo); 4 GB in the paper’s 76 GB Wikipedia setup. Default embedder: Contriever (110M params). It is an index, not an LLM, so no context length applies. Runs on: laptops and workstations. The paper tested an RTX 4090 box with 32 GB RAM and an M1 Ultra Mac. A CPU-only Linux install exists, with no published CPU latency. Performance: matches HNSW downstream accuracy at 90% recall, using under 5% of raw-data storage. Best: 25.5% NQ accuracy, identical to HNSW, using 4 GB instead of 188 GB. Worst: HotpotQA retrieval takes 44.80 s on the Mac, 50.1% of total pipeline time. What is LEANN? LEANN is a low-storage approximate nearest neighbor (ANN) index built on top of FAISS. It targets one problem: vector indexes are often bigger than the data they index. The research paper shows a 76 GB corpus needs 173 GB of embeddings plus a 15 GB HNSW graph. That more than doubles the raw data. LEANN keeps a pruned graph, drops the full vectors, and rebuilds them on demand. The code is MIT licensed and installs with uv pip install leann. How does LEANN recompute embeddings without stalling search? Graph indexes like HNSW visit only a small set of nodes per query. LEANN exploits that. It stores the graph plus a compact product quantization (PQ) table. At query time, it re-embeds only nodes on the search path. 3 techniques keep this practical: Two-level search: cheap PQ distances rank neighbors first. Only top candidates get exact recomputation. Dynamic batching: candidates from several hops are grouped, for example 64 at once, for GPU efficiency. High-degree preserving pruning: most nodes keep few edges, while the top 2% “hub” nodes keep more. In ablations, two-level search gives an average 1.4x speedup. Adding batching raises it to 1.8x. Embedding recomputation still accounts for about 76% of query latency. (function(){var f=document.getElementById('mtp-leann-frame');window.addEventListener('message',function(e){if(e.source===f.contentWindow&&e.data&&e.data.mtpLeannH){f.style.height=e.data.mtpLeannH+'px';}});})(); How does LEANN perform on benchmarks? The research team tested NQ, TriviaQA, GPQA and HotpotQA against a 60M-passage RPJ-Wiki datastore. Latency is measured at 90% Recall@3. All figures below are self-reported in the research paper. On NQ, with Qwen3-4B on an RTX 4090, LEANN and HNSW both reach 25.5% downstream accuracy. LEANN uses 4 GB of storage. HNSW uses 188 GB. BM25 scores 18.3%, and PQ compression scores 17.9%. The cost is search time. LEANN retrieval takes 2.48 s on NQ, versus 0.05 s for HNSW. Generation takes about 20 s, so end-to-end latency rises from 20.95 s to 23.34 s. Across datasets on the RTX 4090, LEANN retrieval ranges from 1.12 s (GPQA) to 7.12 s (HotpotQA). The IVF-based recomputation baseline, inspired by EdgeRAG, is up to 200x slower. LEANN improves exact match by up to 11.8% over BM25. The GitHub repo reports similar savings on personal data. 400K chat messages shrink from 1.8 GB to 64 MB. 38K browser entries shrink from 130 MB to 6.4 MB. LEANN vs other vector indexes Storage, latency and accuracy figures come from the LEANN paper v2, Tables 1 and 2 and Section 7.2 (self-reported). License and design rows link to each project’s own source. Where can developers use LEANN? The repo ships ready apps for documents, Apple Mail, Chrome history, WeChat, iMessage, and ChatGPT or Claude exports. MCP readers pull live data from Slack and Twitter bookmarks. LEANN also runs as a semantic search MCP server for Claude Code. On 30 SWE-Bench Pro tasks from ContextBench, the team reports 24.2% initial relevant-code recall versus 11.4% for BM25. The agent also used 8.4% fewer tokens. HNSW is the default backend, with DiskANN for faster search. Generation works with Ollama, Hugging Face, Anthropic, or any OpenAI-compatible API. It supports Ubuntu, Arch, WSL, macOS and Windows. Some apps, like Apple Mail search, are macOS only. Key Takeaways LEANN stores a pruned graph and recomputes embeddings at query time. 60M chunks: 201 GB down to 6 GB in the repo benchmark. Matches HNSW accuracy (25.5% on NQ) at 90% recall. Search costs seconds: 2.48 s vs 0.05 s for HNSW on NQ. MIT licensed and an MLSys 2026 Best Research Paper. Check out the Paper, GitHub Repo and PyPI package. Also, feel free to follow us on Twitter and don’t forget to join our 150k+ML SubReddit and Subscribe to our Newsletter. Wait! are you on telegram? now you can join us on telegram as well. Need to partner with us for promoting your GitHub Repo OR Hugging Face Page OR Product Release OR Webinar etc.? Connect with us The post Meet LEANN: An Open-Source Vector Index that Shrinks a 60M-Chunk RAG Index from 201 GB to 6 GB appeared first on MarkTechPost.

via MarkTechPost
AI

Liquid AI Releases LFM3: 6 Open-Weight 2B-Class Multimodal Models

Liquid AI has released LFM3, its new generation of Liquid Foundation Models. Liquid AI LFM3 is a family of 6 open-weight, 2B-class models built on one shared backbone. The family covers text, vision, audio, omni and decision tasks. Every model can be fine-tuned on a single GPU. Developers can then deploy on phones, laptops, private servers or the cloud. TL;DR Size: LFM3-1.7B has 1.73B parameters and a 256K context window. LFM3-VL-2B has 1.92B and 128K. LFM3-Omni-2B has 2.2B. Runs on: about 1 GB of memory at 4-bit, so phones and laptops. GGUF, MLX, ONNX, FP8 and NVFP4 checkpoints ship. Performance: LFM3-1.7B averages 49.37 across 29 benchmarks, above Qwen3.5-9B at 47.24 (Liquid’s own eval). Best: IFEval 96.66, and AutomationBench 41.29 vs 15.10 for Qwen3.5-9B. Worst: Humanity’s Last Exam (no tools) 7.17, and RULER 128K 65.56 vs 87.11 for Qwen3.5-9B. What is Liquid AI LFM3? LFM3 is a multimodal model family from Liquid AI that branches 6 models from 1 base. All of them start from LFM3-1.7B-Base. The release includes: LFM3-1.7B: a text model for local agents, post-trained for harnesses like Hermes Agent, Pi and OpenClaw. LFM3-VL-2B: the first LFM vision model with video input, sampled at 2 fps. LFM3-Audio-2B: speech in and out across 18 languages, with tool calling. LFM3-Omni-2B: vision, audio and language in one model, answering in text or speech. d1-2B and d1-Omni-2B: decision models that answer many questions in 1 forward pass. How does the tri-hybrid architecture work? The backbone mixes 3 operators across 24 layers. There are 13 gated short convolutions, 5 sliding-window attention layers and 6 global attention layers. The SWA window is 256 tokens. Convolution kernels grew from 3 to 7. Liquid AI team found this mix with its hardware-in-the-loop architecture search. LFM2 was tuned on 1 phone and 1 laptop. LFM3 was tuned across 14 device and backend combinations. These include a Raspberry Pi 5 and the Qualcomm Hexagon NPU. Other changes include head-wise sigmoid gates on attention, partial RoPE on global layers and sandwich norm. The base model saw 22T tokens. The last 3.5T tokens were agent-heavy mid-training, which also extended context to 256K. Post-training ran in 4 stages. SFT on 1.75T tokens came first, then mixed RL. Next, 6 specialist teachers were distilled into 1 model, followed by agentic RL. How does LFM3-1.7B perform on benchmarks? Liquid AI reports that LFM3-1.7B beats LFM2.5-2.6B with 35% fewer parameters. The bigger story is the agent results. On WebArena-Verified it scores 38.42, against 28.94 for Qwen3.5-9B. On BFCL V4 it scores 71.39, against 62.83. Size still wins on knowledge and agentic coding. GPQA Diamond is 63.37, versus 80.40 for Qwen3.5-9B. SWE-bench Verified is 43.40, versus 51.20. Search narrows the gap: Humanity’s Last Exam rises from 7.17 to 13.90 with tools. window.addEventListener("message",function(e){if(e.data&&e.data.mtpLfm3H){var f=document.getElementById("mtp-lfm3-frame");if(f)f.style.height=e.data.mtpLfm3H+"px";}}); How do LFM3-1.7B and its rivals compare? Sources: Liquid LFM3 model card, MiniCPM5-2B card, Qwen3.5-4B card, Gemma 4 E4B card, LFM license. What about vision, audio, omni and decision models? LFM3-VL-2B averages 71.4 across 62 vision benchmarks. Gemma 4 12B scores 71.7 on the same suite. It leads that comparison on visual grounding (77.3) and GUI understanding (83.1). Video (61.2) and agentic tasks (42.5) still trail larger API models. LFM3-Audio-2B reaches 6.2 English WER and a 3.8 UTMOSv2 TTS score. LFM3-Omni-2B scores 70.1 on Daily-Omni, against 61.3 for Qwen2.5-Omni-7B. d1-2B replaces d1-3B with a third fewer parameters. It improves Decision Bench from 83.6 to 85.2. How can developers deploy LFM3? Weights are on Hugging Face, the Liquid AI Playground and the Liquid API. The docs list GGUF, MLX and ONNX for the text and vision models. Runtimes include llama.cpp, MLX, vLLM, SGLang, Ollama and Transformers. On-device, Liquid AI measured llama.cpp at Q4_0 with 32K context. A full 128K context needs 1.6 GB of KV cache. Qwen3.5-4B needs 4.3 GB. On 1 NVIDIA B300 with SGLang, LFM3-1.7B produces 3.6B tokens per day at 1,024 concurrent requests. Liquid AI also showed specialization. A few hours of SFT and GRPO on 1 A100 lifted wind-turbine defect classification to 67%. Key Takeaways 6 open-weight 2B-class models share 1 tri-hybrid backbone. LFM3-1.7B fits in about 1 GB at 4-bit with 256K context. It beats Qwen3.5-9B on BFCL V4, WebArena-Verified and AutomationBench. Knowledge and agentic coding remain clear weak spots. Free commercial use stops at $10M annual revenue. Check out the technical details, model weights and docs. Also, feel free to follow us on Twitter and don’t forget to join our 150k+ML SubReddit and Subscribe to our Newsletter. Wait! are you on telegram? now you can join us on telegram as well. Need to partner with us for promoting your GitHub Repo OR Hugging Face Page OR Product Release OR Webinar etc.? Connect with us The post Liquid AI Releases LFM3: 6 Open-Weight 2B-Class Multimodal Models appeared first on MarkTechPost.

via MarkTechPost