Frontier Model Weight Programming and Open Weight Updates

The large model field has ushered in multiple important developments, with frontier models achieving direct weight editing, Mistral releasing large-scale open weights, and multiple companies disclosing financial and agent application plans.

The essentials

Frontier models directly edit weights to boost math scores

Research from backgrounds including MIT CSAIL shows that the frontier model Astra directly edits the weights of Qwen3.5-0.8B without training, raising math scores from 7.9% to 48.1%. Independent replication calls it weight space programming with results reaching 20 times.

Mistral Large 4 releases open weights and technical report

The Mistral AI team announced that ML4 weights are officially open, featuring 1.05T parameters and a 1M context. Officials state it has frontier capabilities in coding agents and cybersecurity, and is more efficient and compact than the preview version from three weeks ago.

Source ↗

AMD releases Q3 financials and forecasts high revenue growth

AMD released its Q3 2026 financial report, projecting 2027 revenue growth of over 80%, higher than the 77% modeled by analysts. CEO Lisa Su stated that data center chips will account for 80% of next year's sales.

In brief

Anthropic open-sources verifiable inference scheme

The official team released the cryptography-based FlashProve scheme to verify that answers indeed come from a specified model, with efficiency significantly superior to previous open-source methods.

Google AI Studio deprecates some older models

Gemini 2.5 Pro and 2.5 Flash have been deprecated in Google AI Studio, and API requests will be automatically routed starting December 1.

OpenAI introduces daily task evaluation LifeVal

OpenAI built the LifeVal evaluation to measure AI progress on daily tasks, with GPT-6 Astra achieving a perfect score on some tasks.

Revolut migrates engineers to self-hosted GLM

The company migrated about 2,000 engineers to self-hosted GLM for coding, listing Claude Code and Codex as exceptions.

Source ↗

Long-context state tracking accuracy drops to a low

Allen AI's StateBench test shows that frontier model accuracy on the hardest and longest questions dropped to 13%.