Hey there, This issue features Unsloth Studio and the 4 checks it runs before a downloaded model executes. It also covers Anthropic's Claude Haiku 5.5, Liquid AI's open-weight d1-3B and d1-omni-600M, NVIDIA's PivotOPD, and Perplexity's pplx-embed-v2-late. Then 8 Signals: Karotte, Scale's HSS leaderboard, OpenRouter's decision rankings, Extend Operator-1, PACT, and more.

➤ but wait one more thing: TinyFish launched their Ambassador Program (sponsored).

Featured

▲ Unsloth explains what happens when a trusted model repo changes, and how Unsloth Studio re-checks before it runs. 2 incidents frame it. A compromised LiteLLM release passed a security scanner. An infostealer hid in a Hugging Face repo posing as an OpenAI release, with about 244,000 claimed downloads. Studio answers with 4 checkpoints between download and execution. Remote-code approval is tied to a fingerprint of the code, so any change needs new consent. Hugging Face malware verdicts block flagged weight files at load time. Tools run in OS sandboxes: bubblewrap on Linux, Seatbelt on macOS, MXC on Windows. npm rejects packages under 7 days old. The limits are stated plainly. The code scan is not a sandbox, and approved code runs as the Studio user. The weight gate is not fail-closed when scan metadata is missing. Protections cover Studio and Desktop, not a notebook importing the library. Apache 2.0, beta, Docker or desktop on Mac, Windows, and Linux.

➤ wait what? Nebius Opens 2026 Physical AI Awards: Five $150K Compute Prizes Run with NVIDIA (sponsored).

Agentic AI

▲ Anthropic released Claude Haiku 5.5, a small model with 1M context priced at $0.10 per 1M input tokens. $0.10 input and $0.50 output per 1M tokens for prompts up to 100K. Above 100K, rates rise to $0.50 and $2.50. Batch takes another 50% off. Anthropic estimates about 75% cheaper on average than Haiku 4.5 after the new tokenizer, which counts about 30% more tokens on the same text. 1M context, 128K output, text and image input, adaptive thinking with an effort setting. On OSWorld 2.1 offline it scores 72.4%, against 48.9% for GPT-6 Luna and 15.7% for Haiku 4.5. Terminal-Bench 4.0 is 39.2% to Luna's 16.4%. All scores are Anthropic's. Sonnet 5.5 leads every row.

▲ The web is the one API most agents are missing. Databases, calendars and repos have APIs. The open web mostly doesn't. The TinyFish MCP server gives any MCP client four tools: TinySearch, TinyFetch (full pages as markdown, JavaScript included), TinyBrowser for logins and forms, and TinyAgent for multi-step jobs. Search and Fetch are free. The server is on GitHub. We're racing to 300,000 users this October, with 30% extra on every top-up. [Sponsored]

Decision Models

▲ Liquid AI Releases Open-Weight d1-3B and d1-omni-600M: Multimodal Decision Models With Zero Output Tokens d1 was API-only 9 days ago. Now 2 checkpoints are on Hugging Face. d1-3B takes text and images. d1-omni-600M takes text plus 1 image or 30 s of English audio. Both return probabilities over fixed options in 1 forward pass. d1-3B scores 48.57 on Decision Index v0.2.1, top under 10B. Decider 35B-A3B sits at 47.11, Winnow-12B at 50.02. Liquid ran the scorer itself. 8 ms per question on an RTX 4090 with model.compile, 16 ms without. Knowledge is its weak category at 23.8. No latency figures for the 600M. Day-one llama.cpp, Jetson guides. LFM Open License: free commercial use under $10M revenue.

ML/CV/Data Science/OCR

▲ Perplexity AI Releases pplx-embed-v2-late: A 0.6B Edge Model and a 9B Model Scoring 92.4% on MADQA ColBERT-style: 1 vector per token, 128 dimensions each, text and images and PDF pages in 1 space. The 9B scores 92.4% on MADQA, the 0.6B 90.1%. Mixedbread Agentic Search still leads at 93.4%. BrowseComp+ is 64.0% for the 9B, 8.7 points over the best dense model. Weakest is 61.2% on ViDoRe v3 Markdown. A useful trick: index with the 9B and query with the 0.6B. That beats 0.6B on both sides at the same query cost. Index size grows with document length. Scores are Perplexity's; the technical report is not out. MIT weights, hosted endpoint planned.

Signals worth Catching

▲ [Agentic AI] Preference Model's Karotte is an open framework for RL environments with sandboxed, non-privileged actions, hardened on 1M+ runs.

▲ [OCR] Datalab Introduces OmniExtractBench to Fix Bias and Opacity in Extraction Benchmarks.Sponsored

▲ [ML/CV] Scale's Humanity's Sixth Sense tests intuitive visual reasoning on 522 tasks; GPT-6 Astra leads at 53.6%.

▲ [Agentic AI] Workers IO's skills is an MIT agent skill with 5 modes for writing tests that catch edge-case bugs.

▲ [Agentic AI] OpenRouter's decision rankings show Jev 1.13 at 982M weekly requests, with Liquid's d1 new at 2.87M.

▲ [ML/OCR] Extend's Operator-1 is an extraction agent scoring 99.88% on LongExtractionBench, 5 credits per page.

▲ [Agentic AI] Decagon's PACT is an open consent protocol on A2A and OAuth 2.0 with signed receipts per action.

▲ [Agentic AI] Anthropic's Cyber Verification Program gives vetted defenders and red teams reduced-blocking access in 3 tiers.

Hands on Tutorials with Notebook

▲ The Developer’s Guide to NeMo Guardrails for Enterprise AI Safety

▲ Building Agentic Document Intelligence Pipelines: Creating Scientific Figures with AutoFigure

Until next week,
The Marktechpost Team

Sponsorship Opportunities

Keep Reading

View more