Token-Efficient AI Agent Intelligence. ✨ Smart routing · 🧠 Human-like memory · 🛡️ Sandbox Apache 2.0 · open source
California, USA
Joined April 2026
- Tweets590
- Following8
- Followers346
- Likes77
Pinned Tweet
TokenRhythm API just passed 300K+ developers 🎉
For our 11.11 campaign:
Top up ¥50 (RMB) or more in API credits → get the same amount in bonus credits.
• Bonus credits never expire
• No bonus cap
• Multiple top-ups allowed during the campaign
Your model budget just got a second life.
Limited-time offer. Details in the reply 👇
#TokenRhythm #AI #LLM
The visual embodiment of AI:
“Wait, what am I looking at?”
→
“Ah. A majestic lion.”
#GenerativeAI #AI
Model Twitter: “who’s the king now?”
Me: great — which one gets the task?
Fastest isn’t always best.
Smartest isn’t always cheapest.
The real flex is knowing when to run, when to hand off, and when to bring in a different lion. 🦁
Stop crowning winners. Start building relay teams.
#ModelRouting #AI
Love seeing OpenSquilla in the mix!
Huge props to the team for pushing the boundaries.
Keep building! 🚀
基于千问3.8的125B大模型，在真实任务环里用RSI自己转了一圈，高难榜直接暴涨9分多。🤯
没有人工离线洗数据重训，上一代自己干活、自己查漏补缺，直接产出 NeoHorse-1.5-Flash Preview。
为什么这玩意炸裂：125B 总参只激活 6B（基于 Qwen3.8 Flash 做 RSI 进化）——不追求虚荣 SOTA，专打一次任务动辄上百万 token 的真实 Agent 极致性价比。
硬核评测成绩（纯实测不掺水）：
🎯 ALE 高难榜：平均直接暴涨 +9.23 分（物理科学 +20 点以上 / 工程 +13 以上 / 医疗 +10 以上）
💻 NL to Repo 实测：基座 44.24 ➔ 50.16（+5.92）
🔐 安全认证与加密：35.68 ➔ 48.39（净拉升 +12.71）
🌐 Web 开发与内容获取：51.19 ➔ 61.76（+10.56）
🛠️ 开发工具与测试：34.92 ➔ 44.66（+9.74）
怎么翻越自我诊断盲区、分布坍缩和人工迭代瓶颈这三座大山？
多模型差分诊断：多前沿模型同场作答融合出标尺，动态画出模型"能力缺陷画像"
对抗式环境进化引擎：环境像种群一样繁衍筛选，配合会改需求、会反悔的"数字镜像交互体"高压逼迫
群体智能在线学习：依托自家 OpenSquilla 路由调度真实前沿模型群，拿群体差分当信号 on-policy 边跑任务边进化
让模型在真实生态里当数字牛马自我繁殖迭代🔥 闭源昂贵大模型的蒸馏路线是不是真要走到头了？
Via @opensquilla
🔗 链接：github.com/TokenRhythm/NeoHo… | arxiv.org/html/2609.08183v2 | woshipm.com/ai/6489618.html
#NeoHorse #Qwen #OpenSquilla #AgentAI #LocalLLM
Same prompt. Two takes on Beijing 2008.
One chased the spectacle. The other reconstructed the moment.
There’s no universal “best” model—only the best fit for the intent. That’s the bet behind OpenSquilla.
#ModelRouting #GenerativeAI
At this point, picking an AI model feels like ordering from a menu where every dish has a different price and nobody knows which one's actually better 😭
Gemini 4 Argon vs DeepSeek V4.1 Flash vs GPT-6.1 Terra vs Claude Sonnet 5.5
Per 1M tokens, input / output:
Argon: $2 / $10 (doubles on Dec 1)
Sonnet 5.5: $2 / $10
Terra: $1 / $5
DeepSeek V4.1 Flash: $0.30 / $1.20
Terra costs half of what Argon does and I honestly don’t know what to do with that. Would you pay 2x for a better result?
Same model. Same prompt. Different harness.
86x the cost.
Maybe we've been benchmarking the wrong thing all along.
GPT-6 Luna + DeepSeek Harness: a DISASTER on this test. 🐘💀
Same model via ChatGPT. Same prompt.
DeepSeek Harness: 6.0/10 · 67m · ~$0.86
Codex harness: 7.5/10 · 8m · ~$0.01
8.4× slower. Worse AI-judged score. ~86× pricier.
Follow @LatentOperatorX for more AI battles. ⚔️
This is why model choice matters.
Same task, wildly different outcomes.
5.5 FABLE AND OPUS BUILT A RACING GAME. GPT-6 ASTRA WAS JUST LARPING AS A GAME DEV🤡
I added GPT-6 Astra to my childhood rug test
Fable 5.5 (routing)
· best graphics and physics
· smooth animations everywhere
Opus 5.5
· the fullest map, way more stuff on the rug
· LEGO-style toy cars as obstacles
GPT-6 Astra
· the smallest map of the three
· rough graphics, the rug barely looks like a rug
Waiting for Fable on aimlapi.com
Three years ago, AI couldn't get hands right.
Now we're comparing which model makes better worlds.
What are we even going to compare in 2029? 💀
Fable 5.5 vs Gemini 4 Carbon.
Remember what AI looked like 3 years ago? Barely able to hold a conversation without losing the thread. Hands with seven fingers. We laughed at all of it.
Now I’m sitting here watching a model make this from a single prompt, and I’m not laughing anymore.
We’re moving so fast that I can’t even picture what 2029 looks like. Honestly, it scares me a little.
OpenAI rebuilt part of Git to keep up with Codex-scale work.
That may be the most agentic sentence of the week.
Your AI doesn’t just write code anymore.
It creates enough parallel work to turn source control into a scaling problem.
The bottleneck is shifting from “can the model code?” to “can the stack survive its output?”
#AIAgents #DevTools
The best model isn't always the best choice.
Sometimes the smartest decision is knowing when not to pay more.
Forget benchmarks.
Can it make a Rubik’s cube out of jelly… and still make you want to twist it? 🫠
One model understands materials.
One understands geometry.
One understands color.
One understands the budget.
“Best model” is a boring question.
The better question is: who gets this task?
#ModelRouting #AI
WTF DeepSeek. This Rubik's cube is made of f*cking JELLY and it's the jelliest of all 4 AIs
Same prompt: a Rubik's cube made of JELLY.
DeepSeek V4.1 Flash - $6.36, 1h 0m
Opus 5.5 - $11.10, 1h 23m
GPT-6 Astra - $4.27, 1h 13m
Gemini 4 Argon - $0.55, 32 min
Two weeks ago Astra made plastic and Gemini thought jelly was a color. DeepSeek nailed it.
Which one would you twist first?
We need more AI benchmarks like this.
Forget abstract scores.
Give every model a T-rex, a bicycle, one absurd prompt—and let the personalities show. 🦖
Different models, different weird little strengths.
#AI #GenerativeAI
Opus 5.5 vs Gemini 4 vs 6 Astra: who draws the best T-rex on a bike? 🦖
• Gemini 4 Carbon built him a volcano valley with a glowing sun, a waterfall and mushrooms
• Opus 5.5 drew the biggest T-rex with the most expressive face, plus a countryside with power lines and a windmill
• Gemini 4 Argon gave him a gorgeous purple sunset and legs that bend the wrong way
• 6 Astra gave him a scarf and zero contrast
Carbon wins for me, Opus is a close second.
Who’s your winner?
The $131 refrigerator on rails is taking me out 😭
AI workflows are starting to feel like travel:
same destination, wildly different ticket prices, and somehow someone still booked the most expensive route.
Pick the model for the job—not the fan club.
That’s what OpenSquilla is for.
#ModelRouting #LLMCost
Claude Code really charged $131 to build a train that looks like a refrigerator on rails 😭
Meanwhile Kimi and GPT built an entire cinematic bullet train scene for $19 and $54
Same prompt, 4 AI workflows
The final bill
🥇 Kimi K3.1 : $19 / 5 min
🥈 GPT 6 Astra : $54 / 15 min
🥉 Blender MCP : $87 / 24 min
💀 Claude 3.7 Sonnet : $131 / 38 min
Kimi was almost 7x cheaper and 7.6x faster in this comparison
And somehow the cheapest result has the city lights, reflections and atmosphere you’d actually want in a cinematic render
bro imagine paying $131 two years ago just to invent the world’s fastest microwave
Model fans: “My engine is prettier.”
Production teams: “Cool. What did the lap cost?” 😂
One model for the hero shot.
One for the deadline.
One for the task that needs to be done 1,000 times without setting the budget on fire.
Pick the driver. Let OpenSquilla run the pit wall.
#AIAgents #LLMCost
DeepSeek V4.1 Flash 1017 just embarrassed everyone. 🤯
Same prompt: a working four-stroke engine cutaway.
DeepSeek V4.1 Flash 1017: $0.36 · 1h 54m 16s
Claude Opus 5.5: $21.81 · 2h 00m 13s
GPT-6.1 Terra: $0.43 · 9m 09s
GPT-6 Astra: $7.33 · 54m 33s
Opus got it right. No surprise there.
But DeepSeek? As good as Claude, maybe the most beautiful one of the four… for 36 cents.
That's 60x cheaper than Opus. This is nuts.
Yes, Terra is fast. But costing more than DeepSeek? That's a joke. 😂
Which one wins for you? 👇
Continuous-space language models are evolving fast.
And honestly, the landscape has become hard to keep in your head.
We put together a survey + curated map of the field—from diffusion to flow LMs, and everything in between.
If you’ve ever wondered what comes after token-by-token generation, this is a good place to start.
Paper + repo in the reply 👇
#MachineLearning #GenerativeAI
What’s inside:
• 346 papers mapped
• 133 core systems compared
• A taxonomy across representation, generation dynamics, and decoding
• Head-to-head comparisons with autoregressive and discrete-diffusion baselines
• The open problems that still make serving these models tricky
The interesting question is no longer just “can we generate text without tokens?”
It’s which design choices actually hold up end to end.
Repo: github.com/TokenRhythm/aweso…