🚨 The 276B AI model that beats its 975B teacher — on one GPU

⭐ Featured · Sponsored · Mend.io

Securing AI Agents, MCP Servers & LLM Apps: the practical framework

Agents, MCP integrations, and LLM-powered apps are entering codebases faster than security programs can track them — and their behavior isn’t defined by code alone. This free guide gives AI/ML engineers, platform engineers, and AppSec teams a practical see–fix–protect framework, with checklists and templates to start today.

What’s inside: the agentic AI attack surface map across five layers, from prompt injection to poisoned MCP tools · a 12-point agent and MCP misconfiguration checklist · runtime guardrail architectures, prompt hardening, and policy enforcement · a maturity roadmap aligned to NIST AI RMF, OWASP AIMA, ISO/IEC 42001, and the EU AI Act.

Every release this week splits in two: the thing announced and the thing you can run. Qwen announces 2.4 trillion parameters; the deployable artifact is an API today and a 27B checkpoint later. Cogent announces a frontier cyber model; the deployable piece is the harness around it. AMD open-sources everything except commercial rights. Read each story for the second half.

2.4T

Qwen3.8-Max total parameters

180 GB

Inkling-Small NVFP4 VRAM floor

0.630

Ontology 1 P@10 score

01 · Alibaba Qwen · Frontier MoE

Alibaba Qwen Releases Qwen3.8-Max: A 2.4 Trillion Parameter MoE Model and the Most Capable One in the Qwen Family to Date

A frontier flagship used to mean a closed API and nothing else. Qwen3.8-Max ships as a hosted, OpenAI-compatible endpoint today — and Alibaba says the open weights follow next week, alongside an open Qwen3.8-27B. The 27B, not the 2.4T flagship, is the checkpoint that fits ordinary on-premise hardware.

Context runs to 1M tokens. Pricing is $2.00 input and $6.00 output per 1M tokens, with implicit cache reads at $0.25 — cached input is eight times cheaper than fresh, so prefix stability drives cost more than prompt length. It scores 86.6 on Terminal-Bench 2.1, and against its own predecessor FrontierSWE jumps from 40.7 to 73.5. The activated-parameter count is still undisclosed, so open-weights serving cost cannot yet be modeled.

Read on Marktechpost →  · Announcement  · Model page  · Qwen Studio

02 · Thinking Machines Lab · Open weights

Thinking Machines Lab Releases Inkling-Small: A 276B Total, 12B Active Open Weights Multimodal MoE Model

The student beat the teacher. Inkling-Small is a quarter the size of the 975B Inkling, distilled from it in part — and it comes out ahead on SWE-bench Verified at 80.2% versus 77.6%, on Humanity’s Last Exam at 31.6% versus 29.7%, and on ARC-AGI-2 at 40.1% versus 36.5%. Weights ship under Apache 2.0.

The deployment story is the NVFP4 checkpoint: 180 GB aggregated VRAM on a single B300, against 600 GB for BF16. It reasons natively over text, images and audio with a 1M-token context. The honest trade: factual recall regressed, with SimpleQA Verified falling to 20.6% from Inkling’s 43.9%.

Read on Marktechpost →  · Announcement  · Hugging Face  · Model card

03 · Cogent · Security

Cogent AI Team Releases VR-1: A Frontier Cyber Reasoning Model That Composes and Verifies Enterprise Attack Paths

Most cyber benchmarks reward describing an attack. IntrusionBench, released alongside VR-1, scores nothing for narration — the agent has to execute the chain and produce checkable evidence. VR-1 is post-trained for exactly that: investigating under partial information, composing evidence across cloud, identity, code and CI/CD, and verifying the actual objective.

Cogent reports roughly twice as many proven attack paths at about a quarter of the cost, measured as black-box pass@3 against Kimi K3, Claude Opus 4.8 and GLM-5.2 — and says plainly that harness-matched, the gap nearly closes, and its own black-box success rate stays under 30%. The model is gated to vetted enterprises; the model-agnostic AI Harness is the broadly deployable piece.

Read on Marktechpost →  · Announcement  · Research  · IntrusionBench

Sponsored · Datalab

Marker v2: documents in, Markdown out

PDFs, scans, DOCX and PPTX go in; Markdown, JSON and HTML come out — and tables, inline math and reading order survive the trip. Marker v2 scores 76.0 on olmOCR-bench and sustains 2.9 pages per second on one B200, over 5× MinerU’s pipeline backend while beating Docling on both accuracy and speed.

Fully open source under Apache 2.0. Clone the repo, point it at your document pile, and feed the clean output straight into your RAG or training pipeline.

GitHub repo → · Benchmark breakdown

04 · Onton · Search

Onton Releases Ontology 1: A Neurosymbolic Search Model That is 2.7x More Accurate than the World’s Best E-commerce Search Engines

There is no product filter for “pet-friendly.” Ontology 1 does not trust the seller’s label either — it reasons from properties more likely to be objective, like fiber, weave and construction, in an explicit knowledge graph you can inspect, and flags claims the product data contradicts.

On Subtext-Decor-90, scored by three independent LLM judges, it reaches P@10 0.630 against Google Shopping’s 0.543 and Amazon’s 0.469 — while indexing roughly 1% of their catalogs, winning 52 of 90 queries outright. Judge agreement is modest (Krippendorff’s alpha 0.465), but all three judges rank the engines identically. Live on Onton.com; no public API or open weights.

Read on Marktechpost →  · Research  · Benchmarks  · Dataset

05 · AMD · Fully open MoE

AMD Releases Instella-MoE-16B-A3B: A Fully Open Mixture-of-Experts LLM With 2.8B Active Parameters Trained On Instinct GPUs

Most “open” releases give you a checkpoint. AMD is publishing weights from every training stage, plus data mixtures, training configs and inference code — a complete end-to-end MoE recipe trained from scratch on Instinct MI300X and MI325X. The model holds 16B total parameters and activates 2.8B per token: 2 shared experts plus 6 routed from 64.

Gated MLA and FarSkip-Collective deliver a 12.7% pre-training speedup and up to 39.2% lower time to first token under expert parallelism. The base checkpoint averages 76.7, the strongest among fully open models. The catch: weights ship under a research-only ResearchRAIL license — the MIT-licensed training code is the more reusable asset.

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

06 · Supabase · Agent evals

Supabase Releases Evals: an Open Source Benchmark That Scores Claude Code, Codex and OpenCode on Real Supabase Tasks

Most agent benchmarks run against mocks. Every Supabase Evals scenario boots a real containerized stack — agents call the actual MCP server and CLI, then get scored by deterministic checks plus LLM-as-a-judge, with one retry before grading. It is Apache-2.0 and runs locally via pnpm.

Skills mattered least for top models, most for smaller ones: Sonnet 5 rose from 78% to 100% with skills loaded, GPT-5.6 Sol from 89% to 100%. One finding worth quoting to your team: Codex reads roughly 8 docs pages per scenario versus about 2 for Claude Code, which checks docs in under 40% of scenarios even with skills loaded.

Read on Marktechpost →  · Announcement  · GitHub  · Leaderboard

MiniMax · Video generation

MiniMax Releases MiniMax H3: An Omni-Modal Video Model That Generates 15-Second 2K Clips With Native Stereo Audio

One model reads text, images, video and audio as unified context and returns 2K video with native stereo sound, 4–15 seconds. The H3-VAE tokenizer delivers a stated 4× gain in effective sequence length. API today; open weights promised in the coming days, not yet shipped.

Read on Marktechpost → · Announcement · API docs

DeepSeek · Model update

DeepSeek Upgrades DeepSeek-V4-Flash-0731 with Major Agentic and Coding Gains

Same 284B/13B-active MIT-licensed architecture; the jump is post-training only. Terminal Bench 2.1 moves from 61.8 to 82.7, and API pricing sits at $0.14 per 1M input and $0.28 per 1M output tokens. All benchmark numbers are vendor-reported on an unreleased harness — run your own evals first.

Read on Marktechpost → · Hugging Face · Pricing

One more thing…

Marktechpost · Hands-on tutorial

A Tutorial on GeoAI: Designing Footprint Extraction from NAIP Imagery Using U-Net, Grounding DINO, SAM, and Mask R-CNN

After seven releases where the deployable artifact is an API, a partnership or a license, here is the counterpoint: a complete pipeline where everything runs. Raw NAIP aerial imagery in, cleaned and regularized building-footprint polygons out, end to end in one Colab notebook.

You train a U-Net with a ResNet-34 encoder on 512-pixel georeferenced chips, run sliding-window inference on an unseen scene, orthogonalize the predicted masks into polygons, and score IoU and F1. Then you compare three paradigms on the same scene: your trained model, zero-shot Grounding DINO plus SAM from text prompts, and a pretrained Mask R-CNN — and extend the pipeline to any real area using Microsoft Planetary Computer imagery and Overture Maps labels.

Read on Marktechpost → · GeoAI on GitHub

Sponsor AI Dev Signals

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. Sponsor copy 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 two in action — this issue’s placement for Mend.io → and for Datalab →