▲ Google Cloud just announced Agent Compute Engine, a compute service that gives every AI agent its own hardware-isolated microVM with zero credentials inside. The key idea is splitting each agent into a harness (the brain) and a sandbox (the hands). The sandbox has no network card and no tokens. Traffic leaves over vsock, and Agent Gateway swaps in the real credential off-node, so a prompt-injected agent finds nothing to steal. Idle agents are checkpointed to zero active resources and resumed from snapshots, with copy-on-write memory shared across microVMs. Its open-source sibling, Agent Substrate on GKE, reports 500 activations/s at under 500 ms suspend/resume across 200,000 agents. It is a trade-off though: Google publishes no boot time, density or price for Agent Compute Engine itself, and it is not available to customers yet. It already powers Gemini agents at Google and enters preview in the coming months. Agent Substrate is open source and available on GKE for non-production use today...... Full analysis: marktechpost.com/2026/10/22/… Docs: cloud.google.com/blog/topics… Repo: github.com/agent-substrate/s… @googlecloud @GoogleAI @googleaidevs

▲ You.com just launched Inference Router, an OpenAI-compatible API that runs open-weight models with production guarantees. The key idea is ownership of the stack. Requests run first on You.com's own capacity, then fail over to vetted US partners. Web search runs on the same server as the model, and zero data retention covers both. On You.com's LiveNewsBench run, GLM-5.3 with You.com search scored 86.53%, tying GPT-6.1 Sol at 18% less per question. It is a trade-off though: token prices sit near market, with Kimi K3 at $2.85 / $14.25 vs $3.00 / $15.00 at Together AI. 12 models live, from GPT-OSS 120B to Kimi K3. At least 99.9% uptime SLA, 100% US-hosted, 0% platform fee, pay-as-you-go...... Full analysis: marktechpost.com/2026/10/22/… Launch: you.com/resources/introducin… Models: you.com/product/inference-ro… Docs: you.com/docs/guides/inferenc… @youdotcom

▲ Liquid AI just released RIPL, an inference engine that puts small LLMs directly in the request path of live production systems. The key idea is optimizing the whole path, from request arrival to the finished response, for tail latency under load. Liquid tested the same LFM2.5-350M model on vLLM and on RIPL, and RIPL's p95 stayed lower. So part of the gain is the engine, not only the model. On 1 B200, LFM2.5-1.2B on RIPL hit 34.6 ms p95 vs 498 ms for Qwen3.5-0.8B on vLLM (32 in, 32 out, 1,000 req/s). On 256 in, 128 out, the gap grew to 28.4x. It is a trade-off though: LFM2.5-2.6B reached 76.4 ms on the short workload, above the 50 ms headline, and every number is self-reported. RIPL is available now for enterprise LFM deployments on NVIDIA H100, H200, B200 and RTX PRO 6000. It is closed, serves LFMs only. Shopify already runs LFMs at sub-20 ms in production search...... Full analysis: marktechpost.com/2026/10/22/… Blog: liquid.ai/blog/low-latency Calculator: liquid.ai/low-latency-calcul… Models: huggingface.co/LiquidAI @liquidai @maximelabonne

▲ Gradium just open-sourced Phonon, a 31M-parameter text-to-speech model that runs fully on your device, with no GPU and no cloud call. The key idea is a tiny causal transformer that predicts 1 audio latent per 80 ms frame. A 1-step flow-matching head and a small Mimi decoder turn it into 24 kHz audio, streamed as it generates. The numbers: 1.70% WER on 500 hard sentences (phone numbers, emails, codes) vs 3.50% for Kokoro-82M. First audio in 6.9 ms on a MacBook Air M5 CPU vs 517 ms for Kokoro. It is a trade-off though: on the easier Seed-TTS set, Phonon 31m scores 0.51% WER and Kokoro wins at 0.45%. Phonon 31m is English-only, 64.5 MB, Apache 2.0, and runs 3.7x faster than real time on a Raspberry Pi 5. The 5-language 40m models are CC BY-NC...... Full analysis: marktechpost.com/2026/10/22/… Model: huggingface.co/gradium/phono… Repo: github.com/gradium-ai/xn-ptt… Blog: gradium.ai/blog/release-phon… @GradiumAI

▲ IBM just released Granite Speech 5.0 2B, an Apache 2.0 speech model that transcribes and translates speech in 1 model. The key idea is a CTC Conformer encoder feeding a Granite 4.0 1B LLM through a q-former. Audio reaches the LLM at 10 Hz. Prompts switch between transcription, translation, language ID and keyword lists of up to 2,000 terms. It ranks #2 among open-weight models on the Open ASR Leaderboard at 5.68% average English WER. Qwen3-ASR-1.7B leads at 5.40%. Cohere Transcribe sits at 5.77%. It is a trade-off though: on noisy FFASR audio it scores 14.80% vs Qwen3-ASR's 13.41%, and Japanese WER is 10.74%. 2B params, 6 ASR languages, Apache 2.0. Runs in Transformers, vLLM, llama.cpp and mlx-audio. IBM also updated Turbo CTC (19,870 RTFx, now 7 languages) and 4.1 2B Plus...... Full analysis: marktechpost.com/2026/10/22/… Model: huggingface.co/ibm-granite/g… Blog: huggingface.co/blog/ibm-gran… Paper: arxiv.org/abs/2505.08699 @IBM @IBMwatsonx @IBMData @IBMResearch

▲ Perplexity and NVIDIA just released JOLT, a post-training method where one model is both teacher and student, and the teacher is trained to teach. The key idea is simple. Self-distillation gives the model a verified answer and lets it score its own attempts. But a teacher holding the answer can take shortcuts the student cannot follow. JOLT trains the teacher with outcome reward plus a KL pull toward the student, so its feedback stays within reach. On MATH (Qwen3-0.6B), JOLT matched GRPO's accuracy with 6.5x fewer processed training tokens, and about 13x fewer completion tokens. JOLT+ (adds student reward) beat GRPO by 9.9 points on AppWorld and hit 10.96% pass@1 on Terminal-Bench 2.1 vs 8.25%. It is a trade-off though: without student reward, JOLT trails GRPO on AppWorld (51.92% vs 62.20%). It is a method, tested on Qwen3 0.6B to 30B-A3B and DeepSeek-V2-Lite. Full analysis: marktechpost.com/2026/10/21/… Paper: arxiv.org/abs/2610.10447 Blog: perplexity.ai/hub/blog/a-goo… @perplexity_ai @AravSrinivas

▲ Recursive releases Flint, an Apache 2.0 local coding MoE with 40B total and ~2.5B active params (74.2% SWE-bench Verified) The key idea is sparsity tuned for local hardware. Flint has 39.8B parameters but activates only ~2.5B per token: a router picks 6 of 256 experts plus 1 shared expert. Recursive's Eureka Machine, an automated research system, searched the architecture, the 4-bit quantization recipe and the kernels. Recursive reports 74.2% on SWE-bench Verified vs 73.4% for Qwen3.6-35B-A3B, and 125+ tok/s on an M4 Pro MacBook Pro (150+ on a 64 GB RTX Spark). It is a trade-off though: on Terminal-Bench 2.1 Flint scores 54.9%, while dense Qwen 3.8 hits 73.0%. Weights are Apache 2.0. 4-bit builds take about 23 GB and need 48 GB of memory. rec, Recursive's local engine, targets M4 Pro and M5 Pro/Max...... Full analysis: marktechpost.com/2026/10/21/… Model: huggingface.co/RecursiveSI/f… Repo: github.com/recursive-oss/rec Blog: recursive.com/updates/introd… @Recursive_SI @RichardSocher @CaimingXiong @tydsh

▲ Mistral AI just released Voxtral 3, a full voice stack: 3 transcription models, 1 TTS model and a voice agent SDK. The key idea is 1 streaming pipeline. Realtime Transcribe 3 listens and detects end of turn. An LLM like Mistral Small 4 reasons and calls tools. Voxtral Speech 3 replies in the conversation's tone, with each turn under 1 second. Voxtral Small Transcribe 3 scores 2.1% AA-WER, #3 on Artificial Analysis, ahead of Scribe v2 (2.2%). Diarization hits 18.4% cpWER vs 20.3% for MAI-Transcribe-2. It is a trade-off though: MAI-Transcribe-2 is about 4.7x faster, and the realtime model at 240 ms scores 7.4% WER vs 4.2% for Gemini 3.5 Realtime. Public preview on the Mistral API (EU, US, global). Batch ASR from $0.0035/min, TTS $13 per 1M characters, voice agent API from $0.04/min. No open weights announced...... Full analysis: marktechpost.com/2026/10/20/… Announcement: mistral.ai/news/voxtral-3 Docs: docs.mistral.ai/studio/audio… @MistralAI @arthurmensch

▲ Google, Red Hat and IBM Devs just released llm-d 1.0, turning a fleet of vLLM and SGLang servers into one coordinated inference pool on Kubernetes. The key idea is cache-aware routing. The router tracks which replica already holds each agent's prefix and sends the next step there. When GPU or TPU memory fills, KV cache spills to host RAM, SSD or Managed Lustre instead of being evicted. Google reports Gemini's prefix cache hit rate doubled from 35% to 70%, and a 17,000-node GKE benchmark scaled throughput 2.9x. It is a trade-off though: every number is self-reported, and RL time-slicing still leaves roughly 30% of accelerator time idle. Apache 2.0, CNCF Sandbox, runs on NVIDIA, AMD, Intel GPUs and Google TPUs...... Full analysis: marktechpost.com/2026/10/20/… Repo: github.com/llm-d/llm-d Docs: llm-d.ai Technical details: cloud.google.com/blog/topics… @GoogleAI @googleaidevs @RedHat @RedHat_AI @IBM @IBMData @IBMResearch

▲ Google just open-sourced TorchTPU, a native PyTorch backend that makes Google TPUs an ordinary PyTorch device. The key idea is PrivateUse1 registration. Tensors stay plain PyTorch tensors on torch.device("tpu"), with no JAX runtime. Code runs eager by default, then torch.compile lowers it through StableHLO to XLA. The headline number: TorchTitan reached 52% MFU training Qwen 3 8B, with near-linear scaling up to 8,192 Ironwood TPUs. It is a trade-off though: out-of-box eager took 11.49s per step in Google's fine-tuning case, versus 2.78s after compile, kernels and sharding. TorchTPU is co-developed with Meta, Apache 2.0, v0.2.0 on PyPI, and replaces PyTorch/XLA. vLLM, SGLang and Hugging Face Transformers v5 already run on it...... Full analysis: marktechpost.com/2026/10/20/… Repo: github.com/google-pytorch/to… Paper: arxiv.org/abs/2610.24161 Docs: docs.cloud.google.com/tpu/do… @googlecloud @GoogleAI