Artificial Intelligence AI News

H Company Releases NeoMME: A Family of 260M and 800M Single-Tower Multimodal Encoders That Drop the Vision Tower and Causal Decoder No SigLIP2 tower. No causal decoder. No VLM to repurpose. Here's how it works. 👇 (1) Raw patches, not vision-tower features Images are split into non-overlapping 32×32 RGB patches and projected by a 2-layer MLP trained from scratch. Text enters through a 256-dimensional factorized embedding. Both then share every Transformer layer. → 16,384-token context, enough for two 3840×2160 4K UHD images (2) Masked diffusion, not masked language modeling Pretraining is a discrete masked-diffusion text denoiser. Text-only segments draw a corruption rate from U(0,1). Multimodal segments draw from U(0.30,1), which kills the "guess it from the surrounding words" shortcut. → +38.4 points masked-token accuracy from visible page patches at 90% masking (260M) (3) Trained from scratch on a small budget About 524B packed input tokens, roughly 290B of them text-only. ModernBERT saw around 2T text tokens. They used the NorMuon optimizer to squeeze more out of the smaller budget. → 16 H100s for the 260M run, 32 for the 800M Full analysis: https://www.marktechpost.com/2026/09/06/h-company-releases-neomme-a-family-of-260m-and-800m-single-tower-multimodal-encoders-that-drop-the-vision-tower-and-causal-decoder/ Paper: https://arxiv.org/pdf/2609.01657 Technical details: https://huggingface.co/blog/Hcompany/neomme? HF: https://huggingface.co/collections/Hcompany/neomme

NVIDIA Releases Personal AI Router (PAIR): An Open Inference Router That Turns The RTX, DGX Spark And Mac Boxes You Already Own Into One Local AI Cluster. No new cluster API. No agent harness changes. No prompts leaving your network. Here's how it works. 👇 1. It routes, it doesn't execute PAIR is not a new inference engine. Ollama or LM Studio still runs the model on whichever machine PAIR picks. It takes over the default port each engine uses, so the agent keeps talking to the endpoint it already knows. → Proxies Ollama-compatible, LM Studio-compatible and OpenAI-compatible endpoints → The agent decides what to request, PAIR decides where it runs 2. Discovery and trust mDNS finds nearby machines automatically, or you add a node by IP. Trust is bootstrapped by a six-digit PIN shown on one machine and entered on the other. → All node-to-node traffic is blocked until pairing completes → Paired nodes then communicate over mTLS with generated certificates 3. The eligibility filter A node only becomes a candidate once it can actually serve the request. The scheduler weighs five signals: is the node online and ready, is a supported engine enabled, is the exact model present, what is the current job load, what is GPU utilization. → Models don't need to be identical across nodes — PAIR routes by model location → Loading the same tag on more nodes just widens the eligible pool Full analysis: https://www.marktechpost.com/2026/09/04/nvidia-releases-personal-ai-router-pair-an-open-source-virtual-inference-router-that-distributes-local-ai-requests-across-rtx-dgx-spark-and-mac-nodes/ Repo: https://github.com/NVIDIA/Personal-AI-Router Technical details: https://www.nvidia.com/en-us/ai-on-rtx/personal-ai-router/

Anthropic Released Claude Commerce Agents: An Apache-2.0 Blueprint for Shopping and Merchant Agents Across Retail, Travel, Telecom, and Entertainment. No intent router. No subagent per domain. No custom markup for the UI. Here's how it works. 👇 1. One agent loop, skills for the long tail A commerce session is one tightly coupled conversation across many intents. Every handoff to a subagent is state-lossy — the orchestrator holds the cart, the preferences, the history. → Anthropic reports a single agent with skills beat both the one-big-prompt and the subagent design on quality, often at lower cost and latency 2. Prompt or skill, decided by frequency Loading a skill costs a model turn, so anything the agent needs on most turns goes in the system prompt. Safety rules, brand constraints, and key user facts always sit there. → Roughly a third or more of traffic → system prompt; everything else → skill → 5 skills on the shopping agent, 5 on the merchant agent 3. UI components are tools, not tags Most commerce responses are carousels, itineraries, and seat maps. The model calls present_products or present_itinerary with typed arguments; the server validates and the client renders. Old conversations reload without a custom parser. → The layout lives in the messages array, so "the third one down" resolves 4. Prompt caching carries the cost Caching is prefix-based, so the request is ordered global → session → volatile. A timestamp at the top of the system prompt breaks the cache on every request. → 90–99% hit rate is the range Anthropic says to design for → Cached reads cost a tenth of fresh tokens; cache writes carry a ~1.25x premium Full analysis: https://www.marktechpost.com/2026/09/03/anthropic-released-claude-commerce-agents-an-apache-2-0-blueprint-for-shopping-and-merchant-agents-across-retail-travel-telecom-and-entertainment/ Repo: https://github.com/anthropics/commerce-agents Technical details: https://claude.com/blog/the-anatomy-of-effective-commerce-agents

Meta AI Released Muse Spark 1.3: An Agentic Coding Model Doing the Same Work With ~20% Fewer Tool Calls and ~25% Fewer Tokens Than Muse Spark 1.2. No price increase. No new harness. No open weights either. Here's how it works. 👇 1. Fewer round trips, not just better answers Meta trained 1.3 to take fewer turns where they aren't needed, with less verbosity and a cleaner coding style. → ~20% fewer tool calls and ~25% fewer tokens in Meta's internal engineer comparisons 2. It asks instead of guessing On ambiguous prompts it asks a clarifying question. When it stalls it invokes you. Before consequential actions it confirms. → Better calibration on what counts as irreversible 3. One thread, several workflows Given an open-ended objective, it generates its own context from messy and conflicting sources and patches gaps in its own plan. → Maps an incoming prompt to the right task inside a cluttered thread, whether you're steering or interrupting 4. The numbers (Meta's launch scorecard) → 75.4 on DeepSWE v1.1, ahead of Claude Opus 5 at 74.0 and GPT-5.6 Sol at 72.7 → 88.8 on Terminal-Bench 2.1, tied with GPT-5.6 Sol → 59.4 on SWE-Atlas Codebase QnA → 98.5 and 98.1 on MRCR v2 long-context retrieval, inside a 1,048,576-token window Full analysis: https://marktechpost.com/2026/09/03/meta-ai-released-muse-spark-1-3-an-agentic-coding-model-that-uses-20-fewer-tool-calls-and-25-fewer-tokens-than-muse-spark-1-2/ Technical details: https://research.meta.ai/blog/introducing-muse-spark-1-3

Perplexity Open Sources Lily: A Rust + Metal Inference Engine for Qwen3.6-35B-A3B on Apple Silicon Most "runs locally on your Mac" stacks are a general-purpose runtime pointed at whatever model you downloaded. Perplexity just argued that the generality itself is the bottleneck. They open sourced Lily — the local inference engine behind Hybrid Compute in Perplexity Computer. A Rust runtime with hand-written Metal kernels, built for exactly one model (Qwen3.6-35B-A3B) on exactly one chip family (Apple silicon). Neither PyTorch nor MLX is anywhere in the execution path. Here's what's actually interesting: → 4,156 vs 3,388 prefill tokens/s and 170.0 vs 126.4 decode tokens/s against MLX-LM — mean across ten lengths from 256 to 128K tokens, batch 1, one 40-core / 128 GB M5 Max → Fusing 4-bit dequantization into the grouped GEMM, so the expanded weight array never touches unified memory: +77.4% prefill at a 512-token prompt → Keeping the whole routing sequence — histogram, prefix scan, scatter, block map — inside one GPU command buffer: +89% prefill at 512 tokens → GQA packing, so four query heads share one KV row load: +23.8% decode at 32K context → Fixed-block attention layout above 32K: +40.2% decode at 128K Full analysis: https://www.marktechpost.com/2026/09/02/perplexity-open-sources-lily-a-rust-metal-inference-engine-for-qwen3-6-35b-a3b-on-apple-silicon/ GitHub: https://github.com/perplexityai/pplx-garden/tree/main/lily Technical details: https://www.perplexity.ai/hub/blog/optimizing-on-device-inference-for-apple-silicon

Meta Superintelligence Labs Releases Muse Voice Transcribe: One Real-Time Model for Streaming ASR, Diarization, and Endpointing Most production voice stacks are three systems stitched together. One model transcribes, a second separates speakers, and a detector decides when the user stopped talking. Each hand-off adds latency and a new failure mode. Muse Voice Transcribe, announced by Meta Superintelligence Labs this week, collapses those three jobs into a single autoregressive model. Meta calls it its first real-time audio perception model. It performs streaming ASR, speaker diarization for 20+ speakers, and endpointing in one pass, with no required post-processing.... Full analysis: https://www.marktechpost.com/2026/09/01/meta-superintelligence-labs-releases-muse-voice-transcribe-one-real-time-model-for-streaming-asr-diarization-and-endpointing/ Technical details: https://x.com/AIatMeta/status/2094839236016976028

Perplexity Releases Hybrid Compute on Mac: Cloud Agents Orchestrate Down to a Local Model, Gated On Device Agentic assistants have a structural problem: the context that makes them useful — deal documents, privileged files, client records — is exactly the context users cannot send to a cloud endpoint. This week, Perplexity shipped its answer for Mac. Hybrid compute splits a single Perplexity Computer task between frontier models in the cloud and a compact model on the user’s Mac, with an on-device privacy gate deciding what may cross the boundary. Perplexity also open-sourced the classifier behind that gate.... Full read: https://www.marktechpost.com/2026/09/01/perplexity-releases-hybrid-compute-on-mac-cloud-agents-orchestrate-down-to-a-local-model-gated-on-device/ Technical details: https://www.perplexity.ai/hub/blog/introducing-hybrid-compute-on-mac

Google AI Releases TimesFM-3: A 330M Parameter Zero-Shot Foundation Model For Multivariate Time Series Forecasting TimesFM-3 is a 330 million parameter time series foundation model that forecasts multiple related series in a single forward pass. Every TimesFM checkpoint through 2.5 was univariate: one series, its own history, nothing else. TimesFM-3 is pretrained natively for multivariate forecasting on more than 1 trillion time points, and accepts multiple targets, past covariates, and past-future covariates with no task-specific fine-tuning. It takes the top average rank among pretrained foundation models on GIFT-Eval, fev-bench, and the TIME leaderboard, on both point and probabilistic metrics...... Full analysis: https://www.marktechpost.com/2026/08/31/google-ai-releases-timesfm-3-a-330m-parameter-zero-shot-foundation-model-for-multivariate-time-series-forecasting/ HF Card: https://huggingface.co/google/timesfm-3.0-pytorch Technical details: https://research.google/blog/timesfm-3-a-zero-shot-foundation-model-for-multivariate-forecasting/

I've read a lot of environment-scaling papers this year. This is the first one that doesn't generate anything. Google AI Introduces EnvHarness: A Programmable Layer That Turns Static Agent Environments Into Adaptive Training Worlds It wraps an existing environment through the standard reset/step interface, so the original tasks and human-built verifiers stay in place. An LLM designer writes the wrappers against flaws it finds in the agent's own rollouts. - Apache-2.0, code and reproduction drivers on GitHub - Three components: Stage, Contract, Chain - Stage replays actions to move the episode start state - Contract hooks actions, transitions and observations per step - Chain joins two environments into one episode - EnvRigger loop: observe, diagnose, write, validate - Five benchmarks, four domains, one interface - +9.0 points on held-out ALFWorld tasks - 49.6 vs 55.0 average steps on SWE-bench Verified Full analysis: https://www.marktechpost.com/2026/08/30/google-ai-introduces-envharness-a-programmable-layer-that-turns-static-agent-environments-into-adaptive-training-worlds/ Paper: https://arxiv.org/pdf/2608.19880 GitHub Repo: https://github.com/google-research/envharness

Vercel open-sourced the WebGPU library it built to ship the shaders on vercel.com. vgpu runs the same shader in the browser, in headless Node.js, and in your CI tests. • MIT licensed, on npm at v0.3.1 • .wgsl files import and export like TypeScript modules • Reflection generates bindings, no hand-written declarations • Headless Node.js rendering, Dawn-backed • Deterministic mock adapter for CI snapshot tests • 25 KB gzipped fullscreen effect, budget enforced in CI • CLI, agents.md, llms.txt, hosted read-only MCP endpoint • Examples include ONNX Runtime Web demos on WebGPU Full analysis: https://www.marktechpost.com/2026/08/28/vercel-vgpu-webgpu-library-open-source/ GitHub Repo: https://github.com/vercel-labs/vgpu Join our newsletter for similar product releases: https://www.aidevsignals.com/

Google AI Releases Gemini 3.5 Transcribe: A Speech-to-Text Model Reporting 2.6% Average WER Across 85+ Languages They shipped two models instead of one. gemini-3.5-transcribe-live gives sub-second streaming over WebSockets but no diarization and no word-level timestamps, capped at 10-minute sessions. gemini-3.5-transcribe handles pre-recorded files with diarization, word offsets, and custom vocabulary up to 1,000 terms, at roughly half the cost. Key Takeaways: → Pick the endpoint before the model — latency and diarization are mutually exclusive here, and that shapes your whole pipeline → Smart mode can't be combined with timestamps or diarization — a readable summary and an auditable transcript are now two separate API calls → 4.0% WER streaming, 2.6% batch (per Artificial Analysis), 70% faster finalization than Chirp 3, 85+ languages, roughly $0.005–$0.009/min Full analysis: https://www.marktechpost.com/2026/08/27/google-ai-releases-gemini-3-5-transcribe-a-speech-to-text-model-reporting-2-6-average-wer-across-85-languages/ Technical details: https://www.marktechpost.com/2026/08/27/google-ai-releases-gemini-3-5-transcribe-a-speech-to-text-model-reporting-2-6-average-wer-across-85-languages/

Z.ai Releases GLM-5.3-Flash: A 320B-A18B Natively Multimodal MoE With a 1M-Token Context A 320B-parameter MoE that activates 18B per token — 8 of 288 experts across 45 layers. Roughly 5.6% of the network per forward pass, and the reason a model this size can be served at flash-tier economics. It is also the first GLM model with a hybrid sparse-plus-linear attention stack; the vLLM recipe identifies the layers as KDA linear and NoPE sparse MLA. Z.ai reports ~3× less attention compute and a 4.4× smaller KV cache versus GLM-5.3. If the KV cache figure holds, that is what makes a 1,048,576-token window serveable rather than theoretical. Results: 84.3 on Terminal-Bench 2.1 and 63.4 on DeepSWE v1.1, up from GLM-5.2's 46.2. Vendor-reported, harnesses differ per test. Artificial Analysis ran it independently and scored 57 on the Intelligence Index, at ~49 tokens/sec — strong per dollar, slow in absolute terms..... Full analysis: https://www.marktechpost.com/2026/08/26/z-ai-releases-glm-5-3-flash-a-320b-a18b-natively-multimodal-moe-with-a-1m-token-context/ Model weights: https://huggingface.co/zai-org/GLM-5.3-Flash

Alibaba's Qwen team released Qwen3.8-Flash-Next today. It is an open-weight multimodal MoE and an early preview of the architecture that will underpin Qwen4. Here is what actually matters if you are evaluating it. 1. The parameter math is the headline → 125B backbone + 51B N-gram embedding table + 4B MTP module → 180B total on disk, only 6B active per token → 512 experts, 10 routed + 1 shared active per forward pass 2. Attention is a 3:1 hybrid → Three of every four layers run Gated DeltaNet, which holds a fixed-size recurrent state → The fourth runs Qwen Sparse Attention, scoring micro-blocks rather than individual tokens → QSA budget stays fixed at 512 blocks (2048 tokens) as context grows 3. The benchmarks Qwen published → 62.5 SWE-bench Pro, 81.0 SWE-bench Multilingual, 58.7 DeepSWE 1.1 → 73.9 CoWorkBench, 84.5 AndroidWorld, 95.7 MathVision with code interpreter..... Here is my full analysis: https://www.marktechpost.com/2026/08/26/alibabas-qwen-team-releases-qwen3-8-flash-next-a-125b-multimodal-moe-with-6b-active-parameters-previewing-the-qwen4-architecture/ GitHub Repo: https://github.com/QwenLM/Qwen3.8-Flash-Next?spm=a2ty_o06.30285417.0.0.21fbc921rcDEuy&file=Qwen3.8-Flash-Next HF Model Card: https://huggingface.co/Qwen/Qwen3.8-Flash-Next?spm=a2ty_o06.30285417.0.0.1d73c921FsyOPe&file=Qwen3.8-Flash-Next

Comparison of GPU Neoclouds 2026: CoreWeave, Nebius, Lambda, Crusoe, and Groq (Based on Published Pricing and Contracted Power) Here's what's actually interesting: → The same HGX H100 class is $3.85 at Nebius and $6.16 at CoreWeave — a 60% spread before you touch networking, orchestration or support → Lambda has the cheapest published B200 ($6.69); Nebius is the only one listing B300 on-demand ($7.85); CoreWeave is the only one listing GB200 NVL72 ($10.50/GPU) → CoreWeave is the sole Platinum provider in SemiAnalysis ClusterMAX 2.0, and SemiAnalysis reports it commands a 10–15% per-GPU-hour premium on managed clusters → Nebius's committed capacity clears at $20M+ per MW on Q2 deals and $40M+ on Q3 short-term deals, against a ~$12M 2026 base → Power footprint: CoreWeave 4.2 GW+ contracted; Nebius targeting 5 GW contracted by year-end; Crusoe 4.9 GW contracted; Groq scaling from 54 MW to 200+ MW in 2027; Lambda undisclosed → Crusoe is the only one with AMD MI300X/MI355X on its rate card → Groq licensed its inference tech to NVIDIA for $17B (per NVIDIA's annual report); founder Jonathan Ross and president Sunny Madra joined NVIDIA; Groq then raised $350M at $3.5B and became an NVIDIA Cloud Partner Full analysis and report: https://www.marktechpost.com/2026/08/21/best-gpu-neoclouds-2026/

Liquid AI Releases LFM2.5-DSpark Draft Models That Deliver Up to 3.18x Faster Decoding Without Changing Model Outputs Here's what's actually in the release: 1. Three drafters, one per target model → LFM2.5-1.2B-Instruct, LFM2.5-2.6B, LFM2.5-8B-A1B → Each drafter is ~300M params (295.7M / 327.7M / 327.7M) → 5 attention layers, block size 9, ships no vocab weights 2. The speedups are real but uneven → 3.18x on H100 for 8B-A1B on MATH500 (428 → 1362 tok/s) → 2.87x on an M4 Max for 1.2B-Instruct on HumanEval (136 → 389 tok/s) → 2.67x H100 mean for 2.6B (323 → 864 tok/s) → Same 8B-A1B model drops to 1.29x on GSM8K, same GPU 3. Speedup tracks acceptance rate, not model size → 8B-A1B accepts 8.27 of 10 tokens per step on MATH500 → It accepts 4.02 on GSM8K → That single number explains the 3.18x vs 1.29x gap 4. Output quality does not move → Under greedy decoding, a draft token is kept only if it matches the target's distribution → On rejection, the target's own token takes its place → The emitted sequence is identical to baseline by construction > Full analysis: https://www.marktechpost.com/2026/08/20/liquid-ai-releases-lfm2-5-dspark-draft-models-that-deliver-up-to-3-18x-faster-decoding/ > LiquidAI/LFM2.5-1.2B-Instruct-DSpark: https://huggingface.co/LiquidAI/LFM2.5-1.2B-Instruct-DSpark > LiquidAI/LFM2.5-2.6B-DSpark: https://huggingface.co/LiquidAI/LFM2.5-2.6B-DSpark > LiquidAI/LFM2.5-8B-A1B-DSpark: https://huggingface.co/LiquidAI/LFM2.5-8B-A1B-DSpark Technical details: https://www.liquid.ai/blog/lfm2.5-dspark

Cartesia Ships Sonic-3.6: A Streaming TTS Model That Now Leads Both Artificial Analysis Speech Arenas Key Takeaways: 1. It won both Artificial Analysis boards (as of August 18 2026) → Provider Voice: 1,283 Elo (next closest is 1,240) → Controlled Voice: 1,123 Elo 2. The Controlled Voice result is the real finding That board clones every model onto the same eight reference voices. It strips out the voice-talent advantage and scores the synthesis engine alone. Sonic-3.5 is second on it. ElevenLabs Eleven v3 is third. 3. The architecture Sonic runs on state space models, not transformers. Cartesia states sub-90ms time-to-first-audio — vendor-stated model latency, not a measured round trip. Full analysis: https://www.marktechpost.com/2026/08/18/cartesia-ships-sonic-3-6-a-streaming-tts-model-that-now-leads-both-artificial-analysis-speech-arenas/ Technical details: https://www.cartesia.ai/sonic

DeepSeek AI Releases DeepSeek Harness in Developer Preview: An MIT-Licensed Agent Harness Where Everything is a Plugin Here are some key takeaways: 1. There is no privileged core → Models, tools, skills, sessions, sandboxes, storage, loops, scheduling, and the UI are all Cordis plugins → Any of them can be selected, swapped, or extended in configuration, without editing harness source 2. Four runtime modes, one kernel → Standard, Code, Minimal, Creator — each loads a different default plugin set → Minimal keeps two tools, persistent bash and str_replace_editor, for benchmarking models in a bare environment 3. Every run is traceable → An append-only session log records system prompts, reasoning, tool calls and results, subagent scheduling, and every context injection → Resume, fork, search, and replay all operate on the same event stream Full analysis: https://marktechpost.com/2026/08/17/deepseek-ai-releases-deepseek-harness-in-developer-preview/ Repo: https://github.com/deepseek-ai/deepseek-harness

Google AI Just Released Gemini 3.7 Flash: A Coding and Agent Model at $0.75/1M Input Tokens → Built on 3.6 Flash with algorithmic improvements to the reasoning core. Same 1M context, 64K output, March 2026 cutoff. → The gains concentrate in three places: software engineering, document-heavy knowledge work, and web development. The sharper argument is price. Performance: → FrontierCode 1.1: 43.6% vs 34.4% → DeepSWE v1.1: 65.3% vs 48.6% → WebDev Arena: 1588 Elo vs 1538 → AutomationBench: 30.4% vs 17.0% → GDP.pdf: 34.0% vs 22.0% Full analysis: https://www.marktechpost.com/2026/08/13/google-ai-just-released-gemini-3-7-flash/ Technical details: https://blog.google/innovation-and-ai/models-and-research/gemini-models/introducing-gemini-3-7-flash/