🚨 A 30B at 57.00 on SWE-Bench Verified, ~86% throughput at 1% packet loss, and one 12-second demo → 59% robot success

Presented by NVIDIA  LTX-2.5 — an open weights world model, NVIDIA-accelerated, generating a 10-second clip in 6.8 seconds on-prem. Read the launch →

Every release below removes a unit somebody was being charged for. Granite ships open weights instead of an API key. Perplexity runs the whole harness on hardware you own and charges nothing for local steps. GLiNER2.5 makes a forty-word clause cost the same to find as a two-word name. GEN-1.5 turns tens of thousands of gradient steps into ten. Pipette tells you what a configuration actually costs on the device you ship on, and MetaRoCE stops one dropped packet from stranding a cluster. Six layers of the stack. One move.

57.00

Granite 30B, SWE-Bench Verified

~86%

MetaRoCE throughput at 1% loss

59%

one demo, zero gradient steps

01 · IBM · Open models

IBM Releases Granite 4.2: Bringing Native Reasoning and Agentic RL to Open Enterprise Models

An open enterprise model used to mean an instruction-following assistant you were allowed to self-host. Granite 4.2 is built around reasoning instead. Every size emits a chain of thought before it answers, and the chat template exposes a thinking / low-effort / non-thinking switch. The architecture is deliberately boring — dense, decoder-only, no MoE. The training pipeline is the story.

Three sizes, 3B / 8B / 30B, all Apache 2.0, all pre-trained from scratch on roughly 15 trillion tokens, then post-trained through a multi-stage RL chain. The agentic RL block — SWE, Terminal, Search — runs only on the 8B and 30B, and that one choice explains most of the capability gap across sizes. The 30B reports 57.00 on SWE-Bench Verified and 29.24 on Terminal-Bench 2.1; the 8B, 47.67 and 20.56. IBM also shipped Granite Speech 5.0 Turbo CTC at 470M parameters with no LLM backbone at all.

Read on Marktechpost →  · IBM Research blog  · Hugging Face  · GitHub

02 · Liquid AI · On-device benchmarks

Liquid AI Open-Sources Pipette: A Reproducible Benchmarking Suite That Measures On-Device Models, Quantization, Runtime and Hardware Together

A model card reports quality under server-class, full-precision conditions. Nobody ships that. Pipette changes the unit of measurement from the model to the deployed configuration: model + quantization + runtime + device. Built with Artificial Analysis as an independent methodology validator.

The launch dataset covers five performance metrics across more than 1,000 configurations and 30+ models, at context lengths from 256 to 8,192 tokens, verified on a MacBook Pro M5 Max, an iPhone 17 Pro and a Galaxy S26 Ultra. The comparison that makes the case: at Q4_K_M on the Galaxy, Granite-4.0-H-350M keeps 78.4% of its decode throughput out to 4,096 input tokens while Granite-4.0-350M keeps 33.8%. Same parameter count, opposite answer. One caveat worth reading twice: the quality scores come from H100 reference runs matched to on-device performance, not measured on the phone.

Read on Marktechpost →  · Technical details  · Leaderboard  · GitHub

03 · Perplexity · Agent infrastructure

Perplexity Ships Portable Computer on NVIDIA DGX Spark: Local Harness, OS-Enforced Sandbox, and Zero Per-Token Cost for Local Steps

Most local AI ships a model and leaves you to build the rest. Portable Computer ships the harness, orchestrator, planner, tool router, sandbox and connectors as one packaged system on a DGX Spark. Every task starts on the device. When a step needs the live web or frontier reasoning, the orchestrator stops, runs a PII classifier over the context, shows you exactly what would leave the machine, and waits.

On Perplexity’s own 53-task Local Knowledge Work Bench, Computer with PPLX 27B scored 85.4% against 77.6% for the open-source Pi harness on the identical model. The honest number is Terminal Bench 2.1: 59.6% fully local at effectively zero marginal cost, 73.0% with adviser escalation at roughly $0.415 per rollout, against 82.4% at about $0.65 for Claude Opus 5 alone. Escalation narrows the gap without closing it. The hardware gate is real — a GB10-class box, or an RTX GPU with 24 GB of VRAM, Linux first.

Read on Marktechpost →  · Product page  · NVIDIA local AI blog

Sponsored · NVIDIA

The video production stack now fits on one desk

LTX-2.5 is an open weights world model for video generation, real-time applications and physical AI, optimized by NVIDIA for local inference on RTX GPUs and DGX Spark. Native multishot generation renders a full sequence as one coherent output, holding a character’s look shot to shot. A custom Gemma 4 language backbone sharpens complex multi-subject prompts, and a new diffusion video decoder cuts artifacts in high-motion scenes. One person at a desk can lock a branded character with a quick LoRA fine-tune. No studio, no cloud, no IP leaving the machine.

In LTX’s published image-to-video benchmark, a 10-second clip takes 6.8 seconds on-prem on 2x NVIDIA GB200 and 23.7 seconds through the LTX API, against 52 to 70 seconds for the fastest closed alternatives listed. Weights are on Hugging Face, native in ComfyUI on day one, and free for organizations under $10M in annual recurring revenue. LTX reports more than 33 million downloads across the family.

LTX newsroom → · Open weights · GitHub · Docs

04 · Fastino · Information extraction

Fastino Releases GLiNER2.5: A Boundary-Prediction Architecture That Removes Span Enumeration From Information Extraction

Extraction has been a choice between rigid cheap encoders and flexible expensive LLMs. GLiNER2.5 stops enumerating every candidate span against a width grid and predicts where an entity starts and ends instead. Remove the width axis and the ceiling on entity length goes with it: a forty-word indemnification clause now costs the same to locate as a two-word name.

Three Apache 2.0 checkpoints at 74M, 194M and 287M, all CPU-runnable through pip install gliner2[local], shipping with max_len=4096. Across 16 zero-shot benchmarks the multilingual checkpoint reaches 56.17 overall macro F1 against 56.09 for GLiNER2 — close to flat, with the gain concentrated in XNLI, which jumps to 62.30 from 37.55. Joint entity-relation decoding returns schema-valid graphs by construction, which deletes a post-hoc validation layer. No inference provider hosts the checkpoints, so self-hosting is the deployment path.

Read on Marktechpost →  · Technical blog  · GitHub  · Hugging Face

05 · Meta · AI networking

Meta AI Introduces MetaRoCE: A Clean-Sheet RDMA Transport Built for AI-Scale Ethernet

Standard RoCE assumes the network delivers every frame in order, leans on PFC, and discourages the packet spraying that large multiplane fabrics need. MetaRoCE inverts the assumption. The fabric is treated as lossy; ordering, path selection and recovery move into the NIC. Packets spray across many paths, arrive out of order by design, and write straight to their final memory location — no reorder buffer, no head-of-line blocking.

On a 64-node AMD GPU cluster running RCCL collectives on AMD Pensando NICs, MetaRoCE holds roughly 86% throughput at 1% packet loss and keeps delivering useful bandwidth at 10%, converging rather than collapsing. It asks the switch for two things every switch already has, ECN marking and ECMP, which means it also runs on fabrics you do not control. Not deployable yet: the spec, the compliance suite and libsoftmetaroce are expected at the OCP Global Summit in October 2026.

Read on Marktechpost →  · Meta Engineering

One more thing…

Generalist AI · Robot foundation models

Generalist AI Releases GEN-1.5: A Robot Foundation Model That Learns New Tasks From One 3–12 Second Demo

Teaching a robot a new task has meant collecting data and running tens of thousands of gradient steps. GEN-1.5 takes 3 to 12 seconds of sensorimotor data, dropped into a 30-second context window, and does the task. No gradient updates. No fine-tuning. No task-specific programming. Generalist calls it physical prompting.

Across 10 diverse manipulation tasks, one-shot in-context prompting averaged 59% success (±10% std. dev.) straight from the pretrained model. Ten gradient steps on five minutes of data per task raised that to 83% (±9%). Those ten steps move the weights on held-out tasks by less than 0.15%, which suggests the fine-tune is reconfiguring knowledge the model already has rather than building new representations.

None of it was designed in: no architectural changes, no meta-learning loop, no auxiliary objectives. It emerged from more than eight months of continuous pretraining, the way one-shot prompting emerged in GPT-3. A demonstration recorded entirely in simulation works as a prompt for the real robot, despite pretraining containing no simulation data. The tasks are short-horizon and the company says so plainly. There are no weights, no API and no pricing page — read this as a signal about scaling, not as a product.

Read on Marktechpost → · Announcement thread

Sponsor MARKTECHPOST

Put your product in front of the people who build with it.

This list is engineers, researchers, and founders who read a release note before they read a press release. No banner farms, no interstitials. One sponsor per issue, written in your voice, placed where people are already reading.

Newsletter placements, article sponsorship, product launches, GitHub and Hugging Face repo promotion, and webinars. Tell us the goal and we will send the media kit and available dates.

See one in action — this issue’s placement for NVIDIA →