Liquid's laptop-sized agent model, Fish Audio's 16-voice Drama 3 TTS, and idle compaction in Claude Code

A mixture-of-experts model (only a few expert blocks fire for each token) with 24B total parameters and about 2B active per token. It's text-only and reasoning-only, with a 131K-token context, and it's the largest and last model in the LFM2.5 family. Weights are on Hugging Face and run on GGUF, MLX, vLLM, SGLang and ONNX. With llama.cpp you can start it in one line:

llama-server -hf LiquidAI/LFM2.5-24B-A2B-GGUF:Q4_K_M

It was tuned for tool calling and instruction following, and Liquid reports 27.69 on AutomationBench versus 8.72 for Qwen3.6-35B-A3BVendor claim. At 184 tokens per second on an Apple M5 MaxVendor claim, it's worth testing as the brain of a local agent that never sends company data off the machine. Coding is weaker, though: 51% on SWE-Bench VerifiedVendor claim against Qwen3.6's 66.2%, so it fits workflow automation better than code work.

A text-to-speech model that handles up to 16 speakers in one script. Write stage directions in square brackets, like [whispering] or [laughing], and the delivery changes from that point on. It covers 83 languages and is a 35B mixture-of-experts model with about 3B active per token. Weights are on Hugging Face, code is in fishaudio/fish-speech on GitHub, and it's also in the fish.audio app and API (model: drama-3).

If you make audiobooks, game dialogue or ad mockups with several characters, you can now rough out the voice track without booking actors. Fish Audio says it scored 93.2% on Artificial Analysis' pronunciation robustness test, up from 88.1% for S2.1 ProVendor claim, and that exact strings like codes, paths and emails improved from 79.4% to 86.3%Vendor claim.

In the terminal, a conversation over 200K tokens that sits idle for about 54 minutes now gets compacted before its one-hour prompt cache (the stored prefix that saves recomputing the same context) expires. Change the token threshold with CLAUDE_CODE_IDLE_COMPACT_MIN_TOKENS, or turn it off with "idleCompaction": false in settings.json. A new CLAUDE_CODE_DISABLE_FILE_MENTIONS variable makes @ file mentions pass only the path instead of the contents.

npm install -g @anthropic-ai/claude-code@2.1.299

Come back from lunch to a long session and you no longer pay to reprocess the whole conversation. It also closes a few permission holes: an allow rule for mcp__a no longer covers a server named a__b, macOS sandbox deny rules now hold through symlinks, and plan mode no longer runs writes that a PreToolUse hook approved.