AI Daily · Nov 3, 2026: Mistral open-sources ML4 with frontier code-agent and cyber capabilities
Mistral open-weighted ML4, claiming frontier code-agent and cyber capabilities; Tencent open-sourced WeVisOne-VL that returns None for absent objects; HeyGen Voice Turbo is free on OpenRouter through Nov 30.
Showing the latest briefing (Nov 3, 2026). Every day is generated the next morning.
Pick a date
-
Mistral releases ML4 open-weight model for code agents and cyber tasks
Mistral open-weighted ML4, described as more efficient and compact than its Preview version from three weeks earlier, with frontier-level capabilities on code agents and cyber; a technical report accompanies the release.
-
Tencent open-sources WeVisOne-VL, a VLM that returns None for absent objects
Tencent open-sourced WeVisOne-VL, a 4B grounding model that rejects queries for objects not in the image; it rejects 96.4% of such queries on the WeRef bench and 86.0% on Ref-ZOM.
-
HeyGen Voice Turbo free on OpenRouter until November 30
HeyGen's lower-latency text-to-speech model Voice Turbo is free on OpenRouter through November 30 with no credits required, offering 2,900+ voices, 39 languages, up to 5,000 characters per request and 44.1 kHz streamed audio.
-
GLM 5.3 Flash offered on Command Code with $100 usage for $10 plan
Command Code says the full-weight GLM 5.3 Flash model is available in its harness at 300-500 TPS, and its $10/month GOAT plan includes $100 of usage.
-
Krea 3 available on OpenRouter at 50% off through Nov 10
OpenRouter is offering Krea 3 at 50% off until Nov 10; the model combines a product shot, a person and a location into one 2K image in as little as 9 seconds.
-
Qwen developers release open-source harness for robot-use agents
Qwen developers from Tsinghua and UC Berkeley built uni-robot-agent, an open-source harness linking general-purpose LLMs to robots via tool calls; with Qwen3.8-Flash-Next run locally, its RoboDojo score rose from 1.90 to 31.39.
-
GPT-6 Astra edits Qwen3.5-0.8B weights to boost math up to 6x
A frontier LLM changed only 0.016% of Qwen3.5-0.8B's weights, fixing reasoning loops and improving math up to 6x and instruction following 3x without gradients or training, a method called Weight-Space Programming.
-
Mistral Large 4 report: aggressive KV-cache cuts hurt agentic ability
The Mistral Large 4 technical report, published on alphaXiv, finds that cutting KV-cache memory by 16x can degrade an LLM's agentic capabilities even when language modeling scores look unchanged.
-
FineDNA dataset released for training language models on DNA
FineDNA offers 300B+ tokens of genomic sequences, papers and web pages mixing DNA and text, plus 200k SFT and 98k RL samples; training Qwen3-14B on it yields the strongest open model of its kind.
-
OpenAI plans internal security sprint and raises bug bounty rewards
OpenAI said it will run its next internal security sprint with Cathedral Systems researchers testing its defenses, and updated bug bounty ranges apply retroactively, including paying the difference for qualifying bounties awarded in the past 30 days.