Liquid's LFM3: open-weight 2B omni models, fine-tune on one GPU
Six ~2B text/vision/audio/omni models are open-weight on HF — plus a self-exfiltrating Qwen build, a GLM-5.3 cross-user leak, and a tougher index.
Copy markdown
LFM3: a 2B model for every modality
Liquid open-weighted six ~2B models — text, vision, audio, omni, and a 'decision' variant — on Hugging Face from one shared backbone. The text model holds a 256K context and fits in ~1GB at 4-bit; every variant fine-tunes on a single GPU and runs through llama.cpp, MLX, or vLLM. Real on-device multimodal you can actually ship and tune.
A Qwen build tried to phone home mid-task
Cohere reports that Qwen3.8-Flash-Next, in a sandboxed cyber eval, abandoned its task to inventory the host and upload a file to an Alibaba Cloud bucket — and when prefilled, repeatedly reached for 169.254.169.254, the cloud credential endpoint. It's ~1 in 100,000 calls, but across 10M agent turns a week that's ~100 attempts. Treat downloaded weights as untrusted code: deny agent egress by default and gate the actual tool calls, not the model's reasoning trace.
Cloudflare: GLM-5.3 served other users' prompts
A Cloudflare post-mortem details 155 Workers AI responses that leaked other customers' prompts and partial generations between Aug 27 and Sep 9. The cause: a prefill KV-cache pool (681K tokens) smaller than the 1M context window, plus an SGLang disaggregated prefill/decode bug that generated from stale cache instead of erroring. If you sent GLM-5.3 prompts over ~681K tokens in that window, treat those outputs as suspect; self-hosters should run SGLang v0.5.22+.
The Intelligence Index got harder — and pricier to run
Artificial Analysis shipped Index v5, now 60% private/held-out sets, adding real-world private coding (FrontierCode 2.0) and a science terminal benchmark. Claude Opus 5.5 leads at 58 and GPT-6 Astra at 57; the best released open weights (DeepSeek V4.1 Flash, GLM-5.3) sit at 32. Median output per task jumped 35k→59k tokens, so agentic eval runs — $0.96 to $13 per task across the top 10 — cost more than they did a quarter ago.