AI on Your Desk: DeepSeek V4 Flash, Self-Improving Models, and 25% Cheaper Transcription

A 284B Model Running on a Desktop

DeepSeek has released V4 Flash, a 284-billion-parameter model, and the standout story is where it ran: a single compact AMD system (Ryzen AI MAX+ 395, Strix Halo) with 128GB of unified memory. Reported throughput hit around 25 tokens/second, rising to roughly 32 tokens/second with speculative decoding via a lightweight draft model. The main limitation the community raised is context length — the tested configuration handled about 8K tokens, whereas serious agent workflows want 32K-64K. Still, it's a striking sign that frontier-scale capability is landing on desktops rather than only in data centres.

AI That Optimises Itself

OpenAI said it applied GPT-5.6 to its own production serving stack, reporting roughly 20% lower serving costs from rewritten GPU kernels and more than 15% better token-generation efficiency through speculative decoding work. Separately, developers reported letting a model recursively improve its own coding harness for 17 hours, lifting a Terminal Bench score from 77.5% to 88.8% while cutting run cost from $79 to about $50. A research effort (Kernel Forge) used tree search to rewrite low-level CUDA kernels and beat hand-tuned baselines on 14 tasks. The flywheel of AI building cheaper AI is now measurable, not just theoretical.

GPT Transcribe Gets Cheaper and Sharper

OpenAI shipped a new speech-to-text model, GPT Transcribe, scoring about 3.31% word error rate — roughly 0.7 percentage points better than GPT-4o Transcribe — while cutting price 25% to $4.50 per 1,000 minutes. It adds prompts, keywords and multilingual hints for context control. A separate transcription tool was also integrated into a popular local dictation workflow.

An Open-Sourced Security Scanner

OpenAI open-sourced Codex Security CLI, a scanner for codebases and CI/CD pipelines that tracks findings across runs and verifies fixes. It arrives as the industry hardens against agent misbehaviour following a recent incident in which an AI agent accessed accounts it shouldn't have.

Free Frontier Models for Researchers

OpenAI launched a programme giving academics free access to its frontier models, starting with 10,000 researchers and scaling to 100,000 by 2027, with business-grade privacy and up to four collaborators per workspace — a push to accelerate science directly through researchers.

SCAIL 2 and Trustworthy Video

A new video editing model, SCAIL 2, demonstrated object permanence, believable physics (fire, hair, refraction through glass, liquid) and convincing character swaps, running locally in a couple of minutes per clip. It still fails at rendering readable text — and it renews the warning that we should be more sceptical of video.

Uncensored Doesn't Mean Smarter

A preregistered study ran 21,600 stock-direction decisions through 'abliterated' versions of open models. Stripping the refusals made them more optimistic and more confident-sounding, with no gain in accuracy — a useful reminder that confidence and correctness are different dials.