Sunday, November 22, 2026

Thinking Machines: Inkling 2 Mini brings agentic coding to the laptop — a 20GB open-weights model

🤖 Thinking Machines released Inkling 2 Mini on November 20, 2026: an open-weights model with 26 billion parameters and a file of roughly 20GB that runs on a laptop with 32GB of memory. The weights ship under Apache 2.0 on Hugging Face, which makes the release a direct answer to the users who said Inkling 2 Small didn't fit on their hardware — answered in their own size class. The real claim, though, is not the small file: instead of growing its weights, the model runs the middle of its layers three times per token, so it thinks more while writing fewer tokens.

⏱️ 5 min read

📌 In This Article

  1. 📌 What Inkling 2 Mini is, and where it sits in the family
  2. 📌 The looped transformer: why the layers run three times
  3. 📌 What the benchmarks say, and which numbers are the lab's own
  4. 📌 Where you can run the model
  5. 📌 The research variant: what deepening the loop buys you
  6. 📌 Murat Karakaya Academy verdict
  7. 📌 References

1. What Inkling 2 Mini is, and where it sits in the family

🤖 Mira Murati's Thinking Machines Lab announced the model on November 20, 2026 [1]. Inkling 2 Mini is the smallest member of the family: 26.4B parameters in total, 3.4B of them active per token [1]. The company framed the launch from its own official account with the words "users found Inkling 2 Small (594B parameters) too large to run on their own hardware" — Mini is the direct product of that feedback [5][8].

📐 The family's size ladder today reads:

  • 🔹 Inkling (July 15, 2026): 975B total, 41B active [6]
  • 🔹 Inkling-Small: 276B total, 12B active [6]
  • 🔹 Inkling 2 Small (November 5, 2026): 594B total, 17B active [5]
  • 🔹 Inkling 2 Mini (November 20, 2026): 26.4B total, 3.4B active [1][6]

🖼️ The model takes text and image input and produces text; the context window covers up to 1 million tokens [2].

2. The looped transformer: why the layers run three times

🔁 The architecture rests on a design the lab calls a "looped transformer" [1]. Twenty-four physical layers are arranged in four blocks of six: the first and last block run once per token, while the two middle blocks run three times over the same weights. The result is a 24-layer model that behaves as if it had 48 logical layers [1][3].

⚙️ The rest of the architecture [1][2][3]:

  • 🔹 MoE structure: each token is routed to 8 of 128 experts per layer.
  • 🔹 The model was pruned from 384 experts down to 128 during training, then recovered its capacity on roughly 1 trillion tokens distilled from a larger internal Inkling model [8].
  • 🔹 After pruning came supervised fine-tuning, then a reinforcement-learning (RL) run of 2.2 trillion tokens and 2,000 steps [1].
  • 🔹 Most of the weights were trained — and released — at 4-bit NVFP4 precision, with the last four layers at 8-bit MXFP8; total size roughly 20GB [1].

🧊 So the model was not squeezed down after the fact; it was trained at 4-bit precision from the start. That is why running it locally costs you no quantization penalty on top [7].

3. What the benchmarks say, and which numbers are the lab's own

📊 Start with the numbers the lab itself published: 48.2% on DeepSWE v1.1 and 46.5% on the Vals Index v2.2, which they report as beating every open-weight model under 100B [1]. In the comparison table in the model card — run against Gemma 4 31B, Qwen3.8-27B, Nemotron 3.5 Super VL and Muse Glimmer — they lead in 32 of 48 tests; on the Artificial Analysis Intelligence Index v5 row of that table the score is 20.1, the same as Qwen3.8-27B [2][8].

⚠️ The lab also states an important limit on its own: the loop does not let the weights store more information; knowledge capacity depends mostly on total parameter count. Sure enough, on AA-Omniscience, the factuality measure, its score sits in line with models of its own size class [1]. What the loop does return is efficiency in search and tool use: on BrowseComp, DeepSearchQA and MCP Atlas it leads the open models in its size class [1].

⚡ On the energy side, in Stanford researchers' intelligence-per-joule measurement, Inkling 2 Mini scored 0.703 on an NVIDIA DGX Spark; in the same measurement Qwen3.6-35B-A3B got 0.430 and Gemma-4-26B-A4B 0.159 [1].

4. Where you can run the model

  • 🔹 Open weights on Hugging Face, Apache 2.0 license [2].
  • 🔹 Locally with a single command in Ollama: ollama run inkling-2-mini — a 20GB package, 1M context, text and image input; 32GB or more of memory is recommended on the Apple silicon side [7][2].
  • 🔹 vLLM, SGLang, Unsloth, llama.cpp (GGUF) and MLX support arrived on release day [1][3].
  • 🔹 On the Tinker API: $0.15 per million input tokens ($0.015 from cache), $0.50 output and $0.75 training — exactly half of Inkling 2 Small [4].

✈️ The lab shared two demos as well: researchers ran the model on a MacBook Pro (M5 Pro, 48GB) in airplane mode mid-flight, and it stood up a playable flight simulator in a few minutes [1]. In the other, all 20 characters of Thinkyville, a pixel town, are driven by this same model on a single offline MacBook [1].

5. The research variant: what deepening the loop buys you

🔬 Alongside Mini, Thinking Machines released a research model whose loop count can be varied: TML-d24-varloop-research [1]. Where Mini loops a fixed three times, this variant runs anywhere between one and five passes. In the company's preliminary results, an experiment without chain-of-thought stretched its time horizon from 16 seconds on one pass to 37 seconds on three; on a sequential reasoning test, performance kept improving up to ten loops — twice the number it was trained on [1]. As the loops increase, the model's chain of thought gets shorter but stays readable, and the agent-sabotage tests showed no increase [1]. The lab stresses that it is publishing these models openly so researchers can examine looped architectures for themselves [1].

🎓 Murat Karakaya Academy verdict

🧭 Let's keep the data and our opinion apart. What the data shows: this launch moves the argument from "bigger model" to "more compute per token." The looped architecture is no marketing ornament — the lab backs its claim with a monitorability study and an open research variant. Being able to tune how much a model thinks, independent of how many tokens it writes, and keeping the chain of thought readable while doing it, is a design choice the agent era should take seriously.

⚖️ A hard note on the quality of the numbers: most of the benchmarks on offer are the lab's own measurements. On the independent side, the clearest reading we have is a tie with Qwen3.8-27B on the Artificial Analysis index — not a clean win. Then there is the lab's own admission: the loop adds no knowledge; the model's encyclopedic capacity does not reach past its size class.

🎯 Our opinion: Inkling 2 Mini is the strongest candidate in its class for local agent work today, but the "best in class" claim will only mean something once independent tests multiply. The practical side is already clear: a 32GB laptop can carry a serious agentic coding load, and dropping Tinker to half of Small's price sets a reasonable balance between running locally and calling an API. For developers in Turkey the payoff sits right here: a strong coding agent is no longer a datacenter subscription — it is a laptop.

📚 References

  1. 🔗 thinkingmachines.ai — "Introducing Inkling 2 Mini" · Nov 20, 2026 · accessed Nov 22, 2026
  2. 🔗 huggingface.co — "Inkling 2 Mini model card" · accessed Nov 22, 2026
  3. 🔗 huggingface.co — "Welcome Inkling 2 Mini, Cutting-edge Performance for your Local Setup" · Nov 20, 2026 · accessed Nov 22, 2026
  4. 🔗 tinker-docs.thinkingmachines.ai — "Models & Pricing" · accessed Nov 22, 2026
  5. 🔗 thinkingmachines.ai — "Introducing Inkling 2 Small" · Nov 5, 2026 · accessed Nov 22, 2026
  6. 🔗 thinkingmachines.ai — "Inkling — the model family" · accessed Nov 22, 2026
  7. 🔗 ollama.com — "inkling-2-mini library page" · accessed Nov 22, 2026
  8. 🔗 thinkingmachines.ai — "Inkling 2 Mini Model Card (official)" · accessed Nov 22, 2026

Hashtags: #MuratKarakayaAkademi #Inkling2Mini #ThinkingMachines #LoopedTransformer #AgenticCoding #OpenWeights

🇹🇷 Read this article in Turkish

Murat Karakaya Academy · muratkarakaya.net

Thinking Machines: Inkling 2 Mini ile agentic coding dizüstü bilgisayara iniyor — 20GB model, açık ağırlık

🤖 Thinking Machines, 20 Kasım 2026'da Inkling 2 Mini'yi yayımladı: 26 milyar parametreli, yaklaşık 20GB'lık dosyasıyla 32GB bellekli bir dizüstü bilgisayarda çalışan açık ağırlıklı bir model. Ağırlıklar Hugging Face'te Apache 2.0 lisansıyla dağıtılıyor; böylece laboratuvar, "Inkling 2 Small kendi donanımımıza sığmıyor" diyen kullanıcıların geri bildirimine kendi boyutunda bir modelle cevap vermiş oluyor. Asıl iddia ise boyutun küçüklüğü değil: model, ağırlıklarını büyütmek yerine katmanlarının ortasını her token için üç kez çalıştırıyor, böylece daha az token yazarak daha fazla düşünüyor.

5 dk okuma

Bu Yazıda

  1. Inkling 2 Mini nedir, ailenin neresinde duruyor?
  2. Dönen dönüştürücü: katmanlar neden üç kez çalışıyor?
  3. Ölçümler ne diyor, neresi laboratuvarın kendi sayısı?
  4. Modeli nerede çalıştırabilirsiniz?
  5. Araştırma varyantı: döngüyü derinleştirmek ne işe yarıyor?
  6. Murat Karakaya Akademi Yorumu
  7. Kaynakça

🤖 1. Inkling 2 Mini nedir, ailenin neresinde duruyor?

Mira Murati'nin kurucu ortağı olduğu Thinking Machines Lab, modeli 20 Kasım 2026'da duyurdu [1]. Inkling 2 Mini, ailenin en küçük üyesi: 26,4 milyar toplam parametresi var ve her token başına 3,4 milyar parametre aktif çalışıyor [1]. Şirket, modeli kendi resmî hesabından "kullanıcılar Inkling 2 Small'ı (594 milyar parametre) kendi donanımlarında çalıştırmak için fazla büyük buldu" sözleriyle duyurdu; Mini bu geri bildirimin doğrudan çıktısı [5][8].

Ailenin bugünkü boyut sırası şöyle:

  • Inkling (15 Temmuz 2026): 975B toplam, 41B aktif [6]
  • Inkling-Small: 276B toplam, 12B aktif [6]
  • Inkling 2 Small (5 Kasım 2026): 594B toplam, 17B aktif [5]
  • Inkling 2 Mini (20 Kasım 2026): 26,4B toplam, 3,4B aktif [1][6]

Model metin ve görüntü girdisi alıyor, metin üretiyor; bağlam penceresi 1 milyona kadar token kapsıyor [2].

🔁 2. Dönen dönüştürücü: katmanlar neden üç kez çalışıyor?

Modelin mimarisi "looped transformer" (dönen dönüştürücü) adıyla anlatılan bir kurguya dayanıyor [1]. 24 fiziksel katman, altışarlı dört blok halinde dizilmiş: ilk ve son blok her token için bir kez çalışırken ortadaki iki blok aynı ağırlıklarla üç kez çalışıyor. Böylece 24 katmanlık bir model, 48 mantıksal katmanmış gibi davranıyor [1][3].

Diğer mimari ayrıntılar [1][2][3]:

  • 🔹 MoE yapısı: her token, katman başına 128 uzmandan (expert) 8'ine yönlendiriliyor.
  • 🔹 Model, eğitim sırasında 384 uzmandan 128'e budanmış; ardından daha büyük bir iç Inkling modelinden damıtılan yaklaşık 1 trilyon token ile kapasitesini geri kazanmış [8].
  • 🔹 Budamadan sonra gözetimli ince ayar, sonra 2,2 trilyon token ve 2.000 adımlık bir pekiştirmeli öğrenme (RL) turu koşulmuş [1].
  • 🔹 Ağırlıkların büyük kısmı 4 bit NVFP4, son dört katman 8 bit MXFP8 hassasiyette eğitilmiş ve aynı hassasiyette yayımlanmış; toplam boyut yaklaşık 20GB [1].

Yani model, sonradan sıkıştırılıp küçültülmemiş; baştan 4 bit hassasiyetle eğitilmiş. Yerelde çalıştırırken quantization kaybı yaşamamanızın nedeni bu [7].

📊 3. Ölçümler ne diyor, neresi laboratuvarın kendi sayısı?

Önce laboratuvarın kendi açıkladığı sayılar: DeepSWE v1.1'de %48,2 ve Vals Index v2.2'de %46,5 skorla, 100B altı tüm açık ağırlıklı modelleri geçtiklerini bildiriyorlar [1]. Model kartındaki karşılaştırma tablosunda Gemma 4 31B, Qwen3.8-27B, Nemotron 3.5 Super VL ve Muse Glimmer'a açılan 48 testin 32'sinde öndeler; tablodaki Artificial Analysis Intelligence Index v5 satırında skorları 20,1 — Qwen3.8-27B ile aynı [2][8].

Laboratuvar önemli bir sınırı kendisi de söylüyor: döngü, ağırlıkların daha fazla bilgi depolamasını sağlamıyor; bilgi kapasitesi çoğunlukla toplam parametre sayısına bağlı. Nitekim gerçeklik ölçümü AA-Omniscience'ta skoru, kendi boyutundaki modellerle uyumlu [1]. Karşılığında model arama ve araç kullanımını verimli kullanıyor: BrowseComp, DeepSearchQA ve MCP Atlas'ta kendi boyutundaki açık modeller arasında önde [1].

Enerji tarafında Stanford araştırmacılarının joule başına zekâ ölçümünde Inkling 2 Mini, NVIDIA DGX Spark üzerinde 0,703 almış; aynı ölçümde Qwen3.6-35B-A3B 0,430, Gemma-4-26B-A4B 0,159 [1].

💻 4. Modeli nerede çalıştırabilirsiniz?

  • 🔹 Hugging Face'te açık ağırlık, Apache 2.0 lisansı [2].
  • 🔹 Ollama'da tek komutla yerelde: ollama run inkling-2-mini — 20GB'lık paket, 1M bağlam, metin ve görüntü girdisi; Apple silicon tarafında 32GB ve üzeri bellek öneriliyor [7][2].
  • 🔹 vLLM, SGLang, Unsloth, llama.cpp (GGUF) ve MLX desteği yayım günü geldi [1][3].
  • 🔹 Tinker API'sinde milyon token başına girdi $0,15 (önbellekten $0,015), çıktı $0,50, eğitim $0,75 — Inkling 2 Small'ın tam yarısı [4].

Laboratuvar iki gösteri de paylaştı: araştırmacılar modeli uçakta uçuş modundaki bir MacBook Pro'da (M5 Pro, 48GB) çalıştırdı ve model birkaç dakikada oynanabilir bir uçuş simülatörü kurdu [1]. Bir diğer gösteride, Thinkyville adlı piksel kasabadaki 20 karakterin tamamı, tek bir çevrimdışı MacBook'ta aynı model tarafından yönetiliyor [1].

🔬 5. Araştırma varyantı: döngüyü derinleştirmek ne işe yarıyor?

Thinking Machines, Mini ile birlikte döngü sayısını değiştirebilen bir araştırma modeli de yayımladı: TML-d24-varloop-research [1]. Mini sabit üç tur dönerken bu varyant bir ile beş tur arası çalışabiliyor. Şirketin ön sonuçlarına göre zincir-of-thought kullanılmayan deneyde zaman ufku tek turda 16 saniyeyken üç turda 37 saniyeye çıktı; ardışık akıl yürütme testinde performans, eğitildiği sayının iki katı olan on döngüye kadar iyileşmeye devam etti [1]. Döngü arttıkça modelin düşünce zinciri kısalıyor ama okunabilir kalıyor ve ajan sabotajı testlerinde bir artış görülmedi [1]. Laboratuvar bu modelleri, araştırmacıların döngülü mimarileri inceleyebilmesi için açık yayımladığını özellikle vurguluyor [1].

🎓 Murat Karakaya Akademi Yorumu

Veriler ile kanaatimizi ayıralım. Veriler şunu gösteriyor: bu lansman, "daha büyük model" tartışmasını "token başına daha fazla hesap" tartışmasına taşıyor. Döngülü mimari, bir pazarlama süsü değil; laboratuvar iddiasını izlenebilirlik (monitorability) çalışmasıyla ve açık bir araştırma varyantıyla birlikte destekliyor. Bir modelin ne kadar düşündüğünü, yazdığı token sayısından bağımsız olarak ayarlayabilmek — ve bunu yaparken düşünce zincirini okunur tutmak — ajan çağında ciddiye alınması gereken bir tasarım tercihi.

Sayıların niteliğine dair notumuz sert olmalı: öne sürülen benchmark'ların büyük kısmı laboratuvarın kendi ölçümü. Bağımsız tarafta elimizdeki en net okuma, Artificial Analysis endeksinde Qwen3.8-27B ile başa başlık — net bir geçiş değil. Bir de laboratuvarın kendi itirafı var: döngü bilgi eklemiyor; modelin ansiklopedik kapasitesi boyut sınıfının ötesine geçmiyor.

Kanaatimiz: Inkling 2 Mini, kendi sınıfında yerel ajan işleri için şu anki en güçlü aday, ancak "sınıfının en iyisi" iddiası bağımsız testler çoğalınca anlam kazanacak. Pratik taraf ise bugünden net: 32GB'lık bir dizüstü, ciddi bir agentic coding yükünü kaldırabiliyor; Tinker tarafında Small'un yarı fiyatına inilmesi de yerel çalıştırma ile API arasında makul bir denge kuruyor. Türkiye'deki geliştiriciler için asıl karşılık burada: güçlü bir kod ajanı artık veri merkezi aboneliği değil, bir dizüstü bilgisayar meselesi.

📚 Kaynakça

  1. thinkingmachines.ai — "Introducing Inkling 2 Mini" · 20.11.2026 · erişim: 22.11.2026
  2. huggingface.co — "Inkling 2 Mini model card" · erişim: 22.11.2026
  3. huggingface.co — "Welcome Inkling 2 Mini, Cutting-edge Performance for your Local Setup" · 20.11.2026 · erişim: 22.11.2026
  4. tinker-docs.thinkingmachines.ai — "Models & Pricing" · erişim: 22.11.2026
  5. thinkingmachines.ai — "Introducing Inkling 2 Small" · 05.11.2026 · erişim: 22.11.2026
  6. thinkingmachines.ai — "Inkling — the model family" · erişim: 22.11.2026
  7. ollama.com — "inkling-2-mini library page" · erişim: 22.11.2026
  8. thinkingmachines.ai — "Inkling 2 Mini Model Card (resmî)" · erişim: 22.11.2026

Etiketler: #MuratKarakayaAkademi #Inkling2Mini #ThinkingMachines #LoopedTransformer #AgenticCoding #AcikKaynak

Yayın tarihi: 22 Kasım 2026 · muratkarakaya.net

Murat Karakaya Akademi

English edition: Thinking Machines: Inkling 2 Mini brings agentic coding to the laptop — a 20GB open-weights model

Saturday, November 21, 2026

YouTube AI Weekly Trend Report (November 9-16, 2026)

YouTube AI Weekly Trend Report
2026-11-09 → 2026-11-16
This data was compiled by Murat Karakaya Academy.
27 active channels / 43  |  121 videos  |  2.9M total views

Headlines of the Week

  1. 1

    Anthropic's Fable 5.2 wave owned the week: the 'solved all of software' test ran to 310,000 views while channels argued over self-improvement loops, the cost gap against Opus 5.5, and ways to run the model without guardrails.

  2. 2

    Google shipped a family instead of a single model: Gemini 4 Carbon, Hydrogen and Decision 1 opened the week's second-busiest front, with 'a model for every day' and 'cheap reasoning' leading the pitch.

  3. 3

    The GPU math behind local AI is shifting: DeepSeek V4.1 running on 64 GB of RAM with no graphics card in sight, and Qwen tested on 8 GB of VRAM, forced a fresh answer to the subscription-versus-hardware question.

Themes of the Week

Weekly rhythm — daily video count (bars · left axis) + views (line · right axis)

2418126015 videos1509 Nov21 videos2110 Nov22 videos2211 Nov24 videos2412 Nov17 videos1713 Nov7 videos714 Nov11 videos1115 Nov4 videos416 Nov436K views842K views406K views896K views176K views91K views73K views6K views0448K896K

Theme distribution

Claude Fable 5.2 and the Anthropic wave: 29 (%24)The Gemini 4 family: Carbon, Hydrogen, Decision 1: 16 (%13)OpenAI DevDay: Codex, the ChatGPT platform and security: 16 (%13)Local AI and the hardware math: 24 (%20)Image, video and voice generation: 8 (%7)Coding agents and the future of software: 8 (%7)Enterprise agents: the LangChain ecosystem: 4 (%3)The AI agenda: DeepMind, alignment and industry debates: 10 (%8)Learning week: from GRPO to inference engines: 6 (%5)121videos
Claude Fable 5.2 and the Anthropic wave 29The Gemini 4 family: Carbon, Hydrogen, Decision 1 16OpenAI DevDay: Codex, the ChatGPT platform and security 16Local AI and the hardware math 24Image, video and voice generation 8Coding agents and the future of software 8Enterprise agents: the LangChain ecosystem 4The AI agenda: DeepMind, alignment and industry debates 10Learning week: from GRPO to inference engines 6

Trend timeline — each dot a video · x: publish time · y: log views · color: theme · hover: title

1001K10K100K09 Nov10 Nov11 Nov12 Nov13 Nov14 Nov15 Nov16 NovEdit the JSON, not the image: clean AI edits in three steps - Murat Karakaya Akademi (293)3 Ways to Run an AI Agent on Cloud Run - Prompt Engineering (156)Recursive Harness: Self-Improving Agents Without Training - Prompt Engineering (16K)Gemini Decision 1: You Don’t Need an LLM to Make a Decision - Prompt Engineering (38K)Gemini 4 Hydrogen: Google's New Everyday Model - Prompt Engineering (271K)Gemini 4 Carbon : The New Frontier - Prompt Engineering (47K)Gemini 4 Carbon Is Finally Here, And It's a Beast! - Prompt Engineering (53K)Claude Fable 5.2: The New Frontier from Anthropic! - Prompt Engineering (159K)Ripple-65B Preview: CRAZY 80.5 Terminal-Bench From a 65B Open Model? I Checked the Claims - Prompt Engineer 48 (119)Sketchpad: A 0.8B AI Does 100-Digit Math EXACTLY - I Tested It On My RTX 4060 - Prompt Engineer 48 (144)llama.cpp 1.0 Broke My Code? 36 Functions Just VANISHED (Real Test) - Prompt Engineer 48 (295)Qwen3.8 on 8GB VRAM? Unsloth Dynamic V3.1 Reality Check (RTX 4060 vs 4090) - Prompt Engineer 48 (469)SOV-Transcribe-1: Salesforce Released a WILD Open Transcription Model. I Tested It vs Whisper! - Prompt Engineer 48 (171)HyperQwen: I Ran a 27B AI on One GPU! (Qwen3.8-27B vs Stock vLLM, Real Benchmarks) - Prompt Engineer 48 (524)NVIDIA AgentBaton: Can a 4B Model Finish the Job? I Tested It (WILD Result) - Prompt Engineer 48 (369)Duo v1: An AI World Model That Sees, Hears and Talks! (CRAZY Christmas Test, 1 File Missing) - Prompt Engineer 48 (305)LLM Router 2026.11.0: My Rented GPU Answered ZERO Requests (It Still Cost $1.80) - Prompt Engineer 48 (223)Flide: I Tested The "9.5x Faster" AI Inference Engine - It Hit 6x On My GPU (WILD) - Prompt Engineer 48 (469)Nod AI Agent Is WILD: It Refused to Cancel My Subscription (Honest Review) - Prompt Engineer 48 (405)CLAUDE-MEM Review: It FAILED First, Then Actually Remembered (INSANE Claude Code Memory?) - Prompt Engineer 48 (355)Seedance 2.5 4K Stress Test: Draft at 480p First, Then Where It Broke - Prompt Engineer 48 (232)Scrapling's INSANE 349x Speed? I Tested It (Python Web Scraping) - Prompt Engineer 48 (460)Claude Fable 5.2 + Higgsfield MCP: INSANE 5-Job Test (It Caught Its Own Mistakes) - Prompt Engineer 48 (514)Gemini 4 Carbon is Live! - Sam Witteveen (65K)NVIDIA's Nemotron Bolt 74B-A9B 2 Bit - Sam Witteveen (66K)I Tested Every Uncensored Local AI Model (8GB to 32GB VRAM) - Kai (29K)54GB VRAM For $35 The Cheapest Local AI Hack? - Kai (33K)Deepseek V4.1 Flash Runs With No GPU Just 64GB RAM - Kai (65K)Don't Buy a Mac for AI Until You Watch This - Kai (17K)Apple's Local AI App for Mac Is Coming (Harbor) - Kai (112K)Qwen 4 27b: What to Expect from the next local LLM King? - Kai (61K)NVIDIA's 4B Model That Re-Runs Its Own Layers - Kai (30K)Is This the End of NVIDIA's Memory Chip (NCM) - Kai (25K)Ponytail 5 vs Caveman 4 vs RTK 1: Which Codex Setup Wins? - The Gray Cat (4K)Fable 5.2 said no... so I went uncensored - The AI Automators (15K)You're Only Using Half of Claude - Howfinity (1K)The Google Gemini Feature You Need To Start Using Now - Howfinity (2K)The Claude Shortcut Most People Miss - Howfinity (620)You're Missing This Gemini Feature - Howfinity (333)Stop Using Claude Like This - Howfinity (742)Email Marketing For Beginners Full Tutorial - Howfinity (1K)Who controls access to AI? - Matthew Berman (2K)Anthropic Researcher explains how to train LLMs - Matthew Berman (78K)Cheaper AI can lead to more spending - Matthew Berman (2K)Let AI redesign your work - Matthew Berman (1K)AMA with me and the team! - Matthew Berman (10K)Gemini Carbon, NVIDIA CEO AFFAIR, Fable 5.2, Meta Muse Ember, GPT Image 2.6, Minimax H3: AI NEWS - Matthew Berman (111K)What software provides when AI bypasses the interface - Matthew Berman (4K)AI made Box features obsolete before launch - Matthew Berman (10K)Fable 5.2 just solved software - Matthew Berman (11K)Fable 5.2 solved all software... - Matthew Berman (310K)La Cousinade - OpenAI (4K)6 Underrated Ways to Use ChatGPT - OpenAI (13K)How Mozilla uses GPT-6 Cyber to find and fix Firefox vulnerabilities - OpenAI (17K)How Mozilla strengthens Firefox security with GPT-6 Cyber - OpenAI (6K)Cyber at the Frontier: Introducing Daybreak Red Team & Daybreak Defense Factory - OpenAI (53K)What Codex Unlocks for Holiday Extras - OpenAI (7K)How to request and enable Daybreak Blue - OpenAI (8K)How to enable PR security reviews with Codex Security - OpenAI (3K)ChatGPT Images: Three Ways to Get Started - OpenAI (52K)How to run your first Codex Security scan - OpenAI (40K)Building with GPT-Live and Image Generation | DevDay 2026 - OpenAI (10K)Build What’s Next on the ChatGPT Platform with Plugins | DevDay 2026 - OpenAI (9K)Sites and the Next Generation of Web Apps | DevDay 2026 - OpenAI (32K)What Will You Build With Coding Agents That Don’t Quit? | DevDay 2026 - OpenAI (3K)Harness-as-a-Service: Building with the Agents API | DevDay 2026 - OpenAI (14K)Alibaba Qwen4-Max Is Here! Impressive? I Tested It! Beats Gemini Hydrogen, Carbon ?? - AI WITH Rithesh (206)Anthropic BLOCKED China! Now Open-Source AI Models Are 19 Points Behind! - AI WITH Rithesh (632)Gemini 4 Hydrogen Tested! Cheaper and Better Than Carbon? - AI WITH Rithesh (465)Google Gemini 4 Hydrogen! Fast Everyday Cutting edge Intelligence for All ! - AI WITH Rithesh (299)Gemini 4 Carbon did this - AI WITH Rithesh (497)Google Gemini Decision 1 ! Tested vs Liquid d1-ultra & Jev! - AI WITH Rithesh (492)Gemini 4 Carbon Tested! - AI WITH Rithesh (410)Gemini 4 Carbon - Best Demos from Real Users - AI WITH Rithesh (2K)MiniMax H3-Image - 4K AI Images for $0.035… 52 Tests! Open Weights! - AI WITH Rithesh (626)Claude Fable 5.2 vs Opus 5.5 ! Which is Better? - AI WITH Rithesh (365)Claude Fable 5.2 This - AI WITH Rithesh (2K)Muse Ember - Meta's CHEAP & FAST AI Model for Agents! Tested! - AI WITH Rithesh (878)Claude FABLE 5.2 - #1 AI Model - AI WITH Rithesh (411)Tencent DSpine vs DFlash2! Qwen3.8-27B Gets 2.4× FASTER on Mac - AI WITH Rithesh (3K)ComfyUI KSampler Explained: Seed, Steps, CFG and Denoise - pixaroma (12K)Pruna P-H3 Image in ComfyUI | MiniMax H3 for 4K Images & Edits (Ep42) - pixaroma (31K)Give Me 1 Hour, I'll Make LLM Inference Engines Click Forever - Zachary Huang (12K)Give Me 48 min, I'll Make Integral Click Forever - Zachary Huang (4K)I Trust My AI Agents to Merge Code… But NOT Ship It - Cole Medin (2K)Your Coding Agent Has a MASSIVE Blind Spot (You Need to Fix It) - Cole Medin (2K)Your Complete Guide to Fable 5.2 (When to Use It, Effort Levels, Prompting) - Cole Medin (15K)11 More Tiny Coding Agent Fixes With A Stupid Amount Of Payoff (Learn These Now) - Cole Medin (16K)Mock It With AI Before You Write Code #ai #coding #githubcopilot - Burke Holland (2K)Local AI 301: Quantization - Hugging Face (14K)AI Can Do Everything…Except This - Two Minute Papers (45K)DeepMind’s AI Just Did The Impossible - Two Minute Papers (306K)Is Software Over? - Two Minute Papers (186K)LangSmith: The platform to run, improve, and govern agents - LangChain (3K)LangChain Academy New Course: Building and Deploying Managed Agents - LangChain (4K)Fireside Chat with Erik Bernhardsson, Modal - LangChain (2K)How Workday Builds Agent Prototypes in Under a Day with LangChain - LangChain (5K)Should Agents Say "I Don't Know"? | AirBnB at Interrupt 26 - LangChain (1K)How Airbnb builds Trust for Guests and Hosts using AI Agents with LangGraph - LangChain (5K)06 - Multimodal Spatial Biology: Cross Modality Validation - DigitalSreeni (792)05 - Multimodal Spatial Biology: Tumor vs Non Tumor Comparison - DigitalSreeni (869)What could you build with voice? | Andrew Ng #deeplearningai #aiagents - DeepLearningAI (4K)Tabular Foundation Models for Data Science Agents - DeepLearningAI (5K)Can we just switch AI off? | AI alignment - Julia Turc (24K)Gemini Decision 1: Beginners Guide (Complete Setup) - Mervin Praison (469)I Gave Claude Sheets My Excel File. Check What It Drops - Mervin Praison (668)Claude Sheets Just Took On Excel. I REALLY Tested It - Mervin Praison (876)Fable 5.2 Prompting: I REALLY Tested Anthropic's Advice - Mervin Praison (950)Run Claude Code Locally for FREE with Ollama - Mervin Praison (2K)Seedream 5.5 vs Nano Banana: 15 Edits, What Changed - Mervin Praison (2K)Claude Fable 5.2 vs Opus 5.5: What It REALLY Costs - Mervin Praison (6K)Claude Opus 5.5 + Blender: Zero Skills Needed (Complete Setup) - Mervin Praison (978)Underdog Woof: 100% Local Voice AI in 3 Commands - Mervin Praison (1K)ParseBench-2 - LlamaIndex (351)Watch This Before You Consider Freelancing in Tech - Dave Ebbelaar (9K)Learn Local AI in 30 Minutes - Run LLMs locally for FREE - Dave Ebbelaar (19K)Fable 5.2: The Self-Improvement Loop Keeps Loopin’ … (plus Gemini Carbon) - AI Explained (216K)Building with Gemini in Google Antigravity - Google for Developers (19K)Reinforcement Learning with GRPO, Clearly Explained!!! - StatQuest with Josh Starmer (2K)Linear Discriminant Analysis (LDA): Concepts and Code - StatQuest with Josh Starmer (0)Principle Component Analysis (PCA): Concepts and Code - StatQuest with Josh Starmer (0)

01Claude Fable 5.2 and the Anthropic wave29 videos · 1.1M views

Anthropic's new flagship built the week's widest theme, stretching from 300,000-view mega-tests to 'how do I actually use it' guides. The argument has moved past raw capability: effort levels, cost and self-correcting loops now set the tone of the season.

310K09 NovMatthew BermanFable 5.2 solved all software...216K10 NovAI ExplainedFable 5.2: The Self-Improvement Loop Keeps Loopin’ … (plus Gemini Carbon)159K10 NovPrompt EngineeringClaude Fable 5.2: The New Frontier from Anthropic!111K11 NovMatthew BermanGemini Carbon, NVIDIA CEO AFFAIR, Fable 5.2, Meta Muse Ember, GPT Image 2.6, Minimax H3: AI NEWS78K12 NovMatthew BermanAnthropic Researcher explains how to train LLMs65K10 NovSam WitteveenGemini 4 Carbon is Live!61K11 NovKaiQwen 4 27b: What to Expect from the next local LLM King?24K15 NovJulia TurcCan we just switch AI off? | AI alignment15K12 NovCole MedinYour Complete Guide to Fable 5.2 (When to Use It, Effort Levels, Prompting)15K09 NovThe AI AutomatorsFable 5.2 said no... so I went uncensored11K10 NovMatthew BermanFable 5.2 just solved software6K09 NovMervin PraisonClaude Fable 5.2 vs Opus 5.5: What It REALLY Costs4K10 NovThe Gray CatPonytail 5 vs Caveman 4 vs RTK 1: Which Codex Setup Wins?2K11 NovMervin PraisonRun Claude Code Locally for FREE with Ollama2K10 NovAI WITH RitheshClaude Fable 5.2 This2K11 NovAI WITH RitheshGemini 4 Carbon - Best Demos from Real Users1K13 NovHowfinityYou're Only Using Half of Claude97809 NovMervin PraisonClaude Opus 5.5 + Blender: Zero Skills Needed (Complete Setup)95012 NovMervin PraisonFable 5.2 Prompting: I REALLY Tested Anthropic's Advice87612 NovMervin PraisonClaude Sheets Just Took On Excel. I REALLY Tested It74209 NovHowfinityStop Using Claude Like This66813 NovMervin PraisonI Gave Claude Sheets My Excel File. Check What It Drops63215 NovAI WITH RitheshAnthropic BLOCKED China! Now Open-Source AI Models Are 19 Points Behind!62011 NovHowfinityThe Claude Shortcut Most People Miss51412 NovPrompt Engineer 48Claude Fable 5.2 + Higgsfield MCP: INSANE 5-Job Test (It Caught Its Own Mistakes)41110 NovAI WITH RitheshClaude FABLE 5.2 - #1 AI Model36914 NovPrompt Engineer 48NVIDIA AgentBaton: Can a 4B Model Finish the Job? I Tested It (WILD Result)36510 NovAI WITH RitheshClaude Fable 5.2 vs Opus 5.5 ! Which is Better?35513 NovPrompt Engineer 48CLAUDE-MEM Review: It FAILED First, Then Actually Remembered (INSANE Claude Code Memory?)

02The Gemini 4 family: Carbon, Hydrogen, Decision 116 videos · 452K views

Google opened a whole family this week instead of a single launch. Hydrogen, positioned as Flash's heir, aims squarely at the cheap daily-driver segment, while Decision 1 - which claims a decision moment needs no LLM at all - dragged the efficiency debate into architecture itself.

03OpenAI DevDay: Codex, the ChatGPT platform and security16 videos · 272K views

The DevDay 2026 sessions laid out the plan to turn ChatGPT from an app into a platform - Sites, plugins, an Agents API - while Codex Security and the Daybreak line are shaping OpenAI's cyber-defense story into actual product. Mozilla patching Firefox holes with GPT-6 Cyber is the first hard corporate proof of that claim.

04Local AI and the hardware math24 videos · 433K views

Apple shipping a local Mac app called Harbor, $35 VRAM hacks and fresh quantization guides all point the same way: capability is leaking out of the datacenter and onto desks with 8 to 64 GB. Channels warning that 'this is the video that makes you buy a Mac' show hardware shopping now competing head-on with subscription budgets.

112K12 NovKaiApple's Local AI App for Mac Is Coming (Harbor)66K09 NovSam WitteveenNVIDIA's Nemotron Bolt 74B-A9B 2 Bit65K13 NovKaiDeepseek V4.1 Flash Runs With No GPU Just 64GB RAM33K14 NovKai54GB VRAM For $35 The Cheapest Local AI Hack?30K10 NovKaiNVIDIA's 4B Model That Re-Runs Its Own Layers29K15 NovKaiI Tested Every Uncensored Local AI Model (8GB to 32GB VRAM)25K09 NovKaiIs This the End of NVIDIA's Memory Chip (NCM)19K11 NovDave EbbelaarLearn Local AI in 30 Minutes - Run LLMs locally for FREE17K13 NovKaiDon't Buy a Mac for AI Until You Watch This14K11 NovHugging FaceLocal AI 301: Quantization12K14 NovZachary HuangGive Me 1 Hour, I'll Make LLM Inference Engines Click Forever5K10 NovLangChainHow Workday Builds Agent Prototypes in Under a Day with LangChain3K09 NovAI WITH RitheshTencent DSpine vs DFlash2! Qwen3.8-27B Gets 2.4× FASTER on Mac1K09 NovMervin PraisonUnderdog Woof: 100% Local Voice AI in 3 Commands52415 NovPrompt Engineer 48HyperQwen: I Ran a 27B AI on One GPU! (Qwen3.8-27B vs Stock vLLM, Real Benchmarks)46915 NovPrompt Engineer 48Qwen3.8 on 8GB VRAM? Unsloth Dynamic V3.1 Reality Check (RTX 4060 vs 4090)46914 NovPrompt Engineer 48Flide: I Tested The "9.5x Faster" AI Inference Engine - It Hit 6x On My GPU (WILD)30514 NovPrompt Engineer 48Duo v1: An AI World Model That Sees, Hears and Talks! (CRAZY Christmas Test, 1 File Missing)29515 NovPrompt Engineer 48llama.cpp 1.0 Broke My Code? 36 Functions Just VANISHED (Real Test)22314 NovPrompt Engineer 48LLM Router 2026.11.0: My Rented GPU Answered ZERO Requests (It Still Cost $1.80)14415 NovPrompt Engineer 48Sketchpad: A 0.8B AI Does 100-Digit Math EXACTLY - I Tested It On My RTX 406011915 NovPrompt Engineer 48Ripple-65B Preview: CRAZY 80.5 Terminal-Bench From a 65B Open Model? I Checked the Claims009 NovStatQuest with Josh StarmerLinear Discriminant Analysis (LDA): Concepts and Code009 NovStatQuest with Josh StarmerPrinciple Component Analysis (PCA): Concepts and Code

05Image, video and voice generation8 videos · 49K views

MiniMax H3-Image shipping 4K open weights, the Seedream and Nano Banana shootouts and a steady stream of ComfyUI tutorials say the open-source side of image generation has matured. Andrew Ng's call for voice and a crop of local speech models pull audio into the agent ecosystem as well.

06Coding agents and the future of software8 videos · 215K views

'Is software finished?' climbed to 185,000 views, while the practical videos focused on closing the agent's blind spots: say yes to the merge, no to the deploy. On the research fringe, untrained self-improving agent setups - the recursive harness - hint at the next wave.

07Enterprise agents: the LangChain ecosystem4 videos · 14K views

Agent stories from Airbnb and Workday, both running LangGraph and LangChain in production, document the move from demo to a corporate trust-and-control layer. A 'managed agents' course and LangSmith's governance pitch are products of the same maturation.

08The AI agenda: DeepMind, alignment and industry debates10 videos · 388K views

DeepMind's formal-verification breakthrough and the question 'can we close AI alignment?' formed the thinking side of the week, while videos arguing that cheaper AI means more spending, and that agents now bypass software interfaces, keep re-pricing the economy.

09Learning week: from GRPO to inference engines6 videos · 13K views

Training content - GRPO explainers, LDA and PCA, inference-engine architecture, foundation models on a laptop - mirrors the week's appetite for turning viewers into engineers. That inference-speed videos pull big views says performance engineering is the new reading list.

This report was compiled by Murat Karakaya Academy · Data: YouTube RSS (live, no API key) · window: 2026-11-09 .. 2026-11-16 (UTC, computed at run time) · max 15 videos per channel · Prompt Engineer 48 and the OpenAI feed ran into the 15-video RSS cap, so videos from earlier in the week on those two channels may be missing (partial coverage). · render.py @ 20.11.2026 21:46 UTC

Read this report in Turkish: YouTube AI Haftalık Trend Raporu (9 → 16 Kasım 2026)