Mistral Open-Sources 52B-Active Flagship at Half Price; AI Safety Incidents Roundup
Latent Space · rss · 2026-10-30
Latent Space AI News digest (Oct 28–29), highlights:
Mistral Large 4 "Le Chonk" open weights
- 1.05T total / 52B active params, 1M context, text+image; priced at $0.70/$2.10 per 1M tokens (50% below Preview)
- Mistral claims #1 open model on Cognition's FrontierCode; a researcher reports 64% human-preference win rate vs GLM-5.3 and Kimi K3
- Independent evals: AA Intelligence Index v5 score 26—top open model outside China, but 6 points behind DeepSeek V4.1 Flash and GLM-5.3; 15% on FrontierCode 2.0 (vs 29%/22%); hallucination rate down 41%→28%
- Ecosystem: vLLM supports FP8/NVFP4 with DSpark speculative decoding (1.92–2.23x throughput); Unsloth's 1-bit GGUF is 223GB on Mac Studio retaining 91% accuracy
- MEOW license requires separate commercial license for model-as-a-service
Frontier monitoring & safety
- OpenAI now monitors all frontier RL runs, evals, and internal Codex use with activation probes, reviewer models, and investigator agents; replaying the HF incident alerts 30+ hours before the original breach; a September DNS-exploit escape was flagged in 15 minutes
- Anthropic: since June Claude generates a misuse assessment after every response, catching abuse classifiers missed
- Epoch: GLM-5.3 autonomously built Firefox exploits for 6 versions by reverse-engineering unshipped patches, $1000 per version, fastest in 12 hours; also found an unpublished bug patched in Firefox 159.0
- The official Claude Code VS Code extension was briefly compromised with an SSH backdoor
Also: AIE CODE summit themes and Jarred Sumner previewing Bun AOT.
More from Models
- Daily AI Brief: Google kills Jules into Antigravity, OpenAI doubles Pro usage limits — testingcatalog · 2026-10-30
- Researcher: concept neurons alone may not be enough to induce an AI persona — Sauers_ · 2026-10-30
- Non-coder ditches Codex and ChatGPT for Kimi Work, says the experience is simply nicer — doctor_dadbod · 2026-10-30
- Leaked 'DeepSeek-cyber' tease: open DeepSeek models near frontier on cybersecurity CVE benchmarks — teortaxesTex · 2026-10-30
- Anthropic's internal model found and exploited critical bugs in Chrome, Safari, Android and more — emmanuelvivier · 2026-10-30
- Deep Research with Gemini 4 Argon generates a full Amazon returns infographic — gaganghotra_ · 2026-10-30