keep a subagent only if one of these is true
• one loop measurably degrades from context bloat
• tasks run in parallel and need fresh context you throw away after
• one agent can't pick the right tools, or the prompt has conflicting instructions
and even then, call it as a tool w/ a structured result so the main agent keeps the conversation
ALT Screenshot of Anthropic's article listing when to use subagents, and noting agent architectures often become a copy of the org chart (Conway's law).
@AnthropicAI is now telling teams to fold most subagents back into one agent loop
across several enterprise deployments, one agent w/ skills frequently beat both the one-giant-prompt setup and the subagent setup on quality, often cheaper and faster per task
they shipped a kit for the migration too, here's the triage i pulled out of it
ALT Decision tree from Anthropic's guide: keep a subagent if context bloat, parallel fresh-context tasks, or tool-selection conflicts apply, and call it as a tool; otherwise fold it into one loop as a skill, a tool, or a knowledge base.
run it locally w/ the commands from the card:
brew install llama.cpp
llama-cli -hf LiquidAI/LFM3-1.7B-GGUF:Q4_K_M --temp 0.1 --top-k 50 --repeat-penalty 1.1
the card says you need an LFM3-compatible llama.cpp build, a fresh build from main worked
clip is a real run on a plain linux cpu box, no gpu, 1.1G Q4_K_M file, 39 tok/s on that run, w/ a long think block before the answer!
from liquid's blog:
"six 2B-class models (text, vision, audio, omni, and decision) trained from LFM3-1.7B-Base"
"LFM3-1.7B to outperform our previous LFM2.5-2.6B while using 35% fewer parameters"
from the model card:
"at 4-bit precision, it takes about 1 GB of memory"
→ built for local agents, tool calls, structured outputs
→ the card itself says skip it for agentic coding or complex multi-step reasoning!
@liquidai just dropped lfm3, six open-weight models around 2B params on one backbone!
text, vision, audio, omni and two decision models
▸ liquid says the text model fits in about 1 GB at 4-bit
▸ 256k context on LFM3-1.7B
▸ every model fine-tunes on a single GPU, per the blog
Introducing 𝙻𝙵𝙼𝟹: our new generation of Liquid Foundation Models. With a new tri-hybrid architecture and new multimodal encoders, the LFM3 family is the ideal model suite for specialized intelligence.
> 6 model instances, across text, audio, vision, and omni modalities, generative and decision models.
> 𝙻𝙵𝙼𝟹-𝟷.𝟽𝙱: text, 256k context window.
> 𝙻𝙵𝙼𝟹-𝚅𝙻-𝟸𝙱: vision-language model with video-input support for spatial understanding
> 𝙻𝙵𝙼𝟹-𝙰𝚞𝚍𝚒𝚘-𝟸𝙱: tool calls, voice chat, reasoning in 18 languages.
> 𝙻𝙵𝙼𝟹-𝙾𝚖𝚗𝚒-𝟸𝙱: an omni model combines vision, audio, and language.
> 𝚍𝟷-𝟸𝙱 and 𝚍𝟷-𝙾𝚖𝚗𝚒-𝟸𝙱: decision models with typed outputs and calibrated probabilities in a single pass. Omni adds audio input.
> fine-tune on 1 GPU.
> a new tri-hybrid architecture
> architecture optimized for 14 processor classes across GPUs, CPUs, and NPUs.
> 𝙻𝙵𝙼𝟹-𝚅𝙻-𝟸𝙱 is a beast. It performs competitively with Gemma 4 12B (6x larger), and Gemini 3.5 Flash Lite, and GPT-5.6 Luna, with 1.9B parameters.
> 𝙻𝙵𝙼𝟹-𝟷.𝟽𝙱 is #1 on the Pareto frontier of device intelligence 𝚙𝚒𝚙𝚎𝚝𝚝𝚎.𝚕𝚒𝚚𝚞𝚒𝚍.𝚊𝚒
Every modality for specialized intelligence. Pick the one that fits your use case, customize it, and run it on-device.
1/ 🧵
same ticker, i'd be careful w/ the usdc/usdt comparison though
• @RobinhoodApp's are tokenized debt securities from robinhood assets (jersey), economic exposure only
• @Ondo's say they're fully backed by securities at us broker-dealers, w/ a first-priority security interest for holders
• robinhood's mint + redeem shuts on weekends, ondo's 11 paired ones redeem 24/7
the pool treats them 1:1, the claims underneath aren't the same, and that drift is what stablepair's dynamic fee is there for
ALT Comparison card: NVDArh is a tokenized debt security from Robinhood Assets (Jersey) with economic exposure only and weekday mint/redeem; NVDAon is backed by securities at US broker-dealers with a first-priority security interest and 24/7 mint/redeem.
One of the largest tokenized stock issuers is now on Robinhood Chain, with Uniswap as its liquidity layer
@Ondo Stocks and Robinhood Stock Tokens are the same shares, pegged 1:1 like USDC/USDT
That’s what StablePair Hook was built for, now in its first use beyond stablecoins
you don't need 8 GPUs to read it, grab the config + card w/ the @huggingface cli:
pip install -U huggingface_hub
hf download inclusionAI/Ling-3.1-flash-1019 config.json README.md --local-dir ling
to serve it, the card's recommended path is sglang on 8 × H200:
docker pull lmsysorg/sglang:dev-ling-3.1-flash
heads up, the card's sglang example still points at inclusionAI/Ling-3.1-flash, so i'd swap the model path to -1019!
what the model card says:
"560B total parameters, 25B active parameters per token, and a 1M-token context window"
"AA-Briefcase score rises from approximately 1400 to 1524 Elo"
what @ArtificialAnlys says today:
"scores 29 on the Artificial Analysis Intelligence Index"
→ $0.30 per 1M input, $0.90 per 1M output on inclusionAI's API
→ MIT license, open weights!
ant ling just open-sourced ling-3.1-flash-1019, and dropped the base model too!
@AntLingAGI says it scores 44 on the artificial analysis index v4.3.2, up from 41
the live @ArtificialAnlys page shows 29, on their newer v5.0 index
same model, two scoreboards, so always check which index version a launch chart is quoting!
setup if you wanna try it tonight:
1. brew install --cask nativ, or grab the dmg from github releases
2. pick a model, nativ recommends one for your hardware
3. point claude code or codex at it, nativ handles the endpoint
haven't run it myself yet, so all of this is neywa's own claims, not my benchmark!
