Sun, Nov 22, 2026Sunday, November 22, 2026 · 18 stories · 7 min read

Local coding agents get a lift from Thinking Machines' Inkling 2 Mini

Ramit KoulFounder, Software Engineer & Innovator · Published 5:03 AM ET

Thinking Machines released Inkling 2 Mini, an open-weights looped model that runs on a 32GB Mac and posts strong agentic coding scores.1, 2 OpenAI also cut the price of idle voice time by adding a standby mode alongside GPT-Live-1.1 in its API.3 Both releases lower the cost of running agents, and they arrive after a week in which the Nasdaq had its worst run since 2020 and Anthropic's new shares slid back toward their offering price.4 Here are the 2 stories that matter most, then 16 briefs. Numbers in the text link to the references at the end.

Models

Thinking Machines releases Inkling 2 Mini, a looped open model built to run on laptops

Image: VentureBeat

Thinking Machines Lab released Inkling 2 Mini on November 20. It is an open-weights model that the company says is small enough to run on a laptop.2, 1 The model takes text and images, outputs text, and ships under Apache 2.0. It has 26.4 billion total parameters, 3.4 billion of them active, and a checkpoint of about 20GB, which fits a Mac with 32GB of memory, a single DGX Spark or a GPU with 32GB of VRAM.5, 6

The model stores 24 layers. Its middle 12 run three times per token with shared weights, which gives 48 layer applications and 5.4 billion effective active parameters without a larger file.1 Thinking Machines says a baseline without looping needed 140% more inference compute to match the looped model's loss on 512K-token agentic traces.5 The weights ship in NVFP4 and MXFP8, the same formats used in training, so there is no loss from quantizing after training.1

On the company's own benchmarks, the model scores 79.2% on SWE-bench Verified (bash only), 77.9% on SWE-bench Pro v2 Hard and 48.2% on DeepSWE v1.1.5, 6 Thinking Machines also lists weaknesses: it trails denser peers on factual recall, it scores 50.0% on the MultiChallenge multi-turn chat test, and it can fall into repetitive debugging loops.1

The weights are on Hugging Face, along with a research variant trained to loop one to five times. The model is also on Tinker for inference and fine-tuning, and Macs can run it through Ollama.2, 1 vLLM, SGLang and llama.cpp GGUF builds are supported as well.5

Why it matters for builders

A 20GB checkpoint with strong agentic coding scores makes offline coding agents, or several persistent agents, practical on a single workstation.5, 2 Because its factual recall is weak, plan to pair it with search or retrieval tools, which is how Thinking Machines suggests closing the knowledge gap.2, 1

Models

OpenAI brings GPT-Live-1.1 to its API and adds cheaper standby billing for idle voice sessions

OpenAI made GPT-Live-1.1 available in its API under the model ID gpt-live-1.1. This is the full-duplex voice model that has powered ChatGPT Plus and Pro voice since October 15.3, 7 OpenAI says it keeps accents more consistent across languages, handles letters and numbers better and completes tasks more reliably, but it published no benchmark scores for the release.3, 7

The new standby setting works with both GPT-Live-1 and GPT-Live-1.1. After a silence of 5 to 60 seconds, set by the developer, the model pauses while the WebRTC call stays open, and it resumes when someone speaks.3 Standby costs $0.02 a minute, compared with $0.05 a minute for active time.3 The session keeps its ID and history through the pause, though OpenAI warns that the first reply after a short utterance may be slower.7

The voice model can hand reasoning and tool calls to a backend model such as GPT-6 Astra, or to a third-party backend, and that usage is billed separately.7 GPT-Live-1.1 does not support structured outputs, fine-tuning, image or video input. Switching models requires a new session, and the two GPT-Live versions have separate limits on concurrent sessions.7

Why it matters for builders: For voice agents that keep calls open through long silences, standby cuts the cost of idle minutes by 60%.7 Budget for backend reasoning and tool calls on top of voice minutes, since they are billed separately and can cost more than the voice layer.7

In brief

  • Models

    A demo shows GPT-6.2 Astra turning a sketch into a simulated pendulum clock. In a demo covered by Geeky Gadgets, GPT-6.2 Astra modeled every clock part in CAD, wrote its own Python solver for the motion and built a moving Blender scene, though the clock has not been printed or physically tested.8 OpenAI released the model on November 17 as its frontier model for demanding professional work, and says it costs less than its predecessor.9, 8

  • Policy & legal

    China's data regulator plans standards for metering, billing and grading AI tokens. Liu Liehong, director of the National Data Administration, said the agency will improve standards for how AI tokens are measured, billed and checked for quality, and will work to prevent low-quality duplication as so-called token factories spread.10, 11 The agency has not said when the standards will be published.10

  • Industry

    Neoclouds that rent out Nvidia chips face a harder market. Business Insider reports that neoclouds lost their biggest champion when Jensen Huang was ousted from Nvidia, and that Washington then cut off a major customer base.12

  • Research

    OpenAI says its agents now finish nearly half of two- to four-hour research tasks without help. In its report "Research acceleration: The jagged frontier," OpenAI says its agents completed 46% of these tasks alone in October, up from 23% in March, and that the median researcher used about $2,400 of agent time a day at API list prices.13, 14 Success rates vary widely by task, from 79% for monitoring training runs to 25% for writing inference and serving code.13

  • Industry

    Nvidia's strong quarter stands in contrast to a campaign to bring back Jensen Huang. Nvidia beat expectations and says Huang will not return, even as Elon Musk and others circulate a petition for his reinstatement and Huang says he is prepared to come back.15 Interim CEO Colette Kress also said Nvidia will no longer sell H200 chips in China.15

  • Industry

    Large US school districts push back on Google's plan to tie Search to AI Overviews. Google will switch its AI tools off by default for K-12 student accounts early next year, but schools that want Google Search will have to accept AI Overviews.16 Chicago says it may move to other search options, and New York City and Los Angeles have asked Google for a way to turn off the summaries.16

  • Funding & deals

    Anthropic shares give back most of their first-day gain during the Nasdaq's worst week since 2020. Anthropic opened at $111.50 on Thursday and closed Friday at $88.91, a fully diluted market value of $1.6 trillion, behind only six US companies.17 Over the same week the Nasdaq fell 10% and Nvidia dropped 23%.17

  • Industry

    A steep market sell-off deepens fears that the AI trade is unraveling. NPR reports that the S&P 500 and the Dow entered a correction on Friday, as investors weighed higher interest rates, Democratic control of Congress and the leadership turmoil at Nvidia.4 Analysts also point to circular financing: Nvidia invests in and lends to companies that then buy its chips.4

  • Funding & deals

    Google will pay SpaceX $920 million a month for GPU compute. A SpaceX filing shows the deal runs from October 2026 to June 2029, about $30.4 billion over 33 months, for access to roughly 110,000 Nvidia GPUs and related infrastructure.18

  • Research

    Tufa Labs compresses 1GB of Wikipedia to 105.39MB using only a Transformer. Tufa Labs' tufazip uses a 163-million-parameter Transformer as its only predictor and is the best entry on the Large Text Compression Benchmark that relies solely on a neural network.19, 20 It is a research project rather than a product: compression takes about 7.3 days and decompression about 7.5 days on two Nvidia B200 GPUs.19

  • Infrastructure

    Phison shows an SSD design that runs AI frameworks on the drive itself. At SC26, Phison detailed a drive design with a built-in NPU and 64GB of 130 GB/s memory that supports PyTorch, vLLM, OpenVINO and llama.cpp directly on the drive.21, 22 Phison claims it could cut data transfers by 10 to 100 times for suitable workloads, but the design has no product name, sampling date or TOPS rating yet.22

  • Products

    Early iOS 28 rumors point to a more agent-like Siri in 2027. Geeky Gadgets, citing a ZONEofTECH video, reports that CEO John Ternus is pushing for Siri to handle multi-step tasks more autonomously in iOS 28, codenamed Bell, ahead of a rumored WWDC 2027 preview.23 Apple has not confirmed any of the features.23

  • Research

    Columbia and CertiK's TensorLean finds 38 bugs in widely used GPU kernels. The Lean-based verifier checked 149 compiled kernel instances, including every kernel in FlashAttention-2, -3 and -4 and in DeepGEMM, proved 83 correct, and turned up 38 bugs, 23 of them in FlashAttention.24 One was a FlashAttention-4 data race that never showed up in 17,760 test runs.24

  • Research

    A TestMu AI survey finds most teams have merged AI-written code they did not fully understand. TestMu AI's State of AI in Quality Engineering 2026 report found that 81% of teams have done so, and only 41% run any automated check on AI-written code before merging.25, 26 Teams that use one model to grade another did this regularly 29% of the time, against 15% for other teams, and only 13% have an incident response plan for AI errors.25

  • Industry

    China's tech giants double AI spending while their labs earn back only a small share. Rhodium Group expects AI infrastructure spending at 13 listed Chinese firms to rise 103% this year to 932 billion yuan ($139 billion), with revenue from their AI labs covering less than 8% of it.27 DeepSeek chief executive Liang Wenfeng said in a leaked transcript of a call with investors that the company cannot afford to train the largest models.27

  • Industry

    The Wall Street Journal says Anthropic's debut succeeded but failed to revive the AI trade. The newsletter notes that Anthropic shares closed up 17% on their first day while the broader market slumped.28 One Journal columnist argues that profits may shift toward companies that apply AI rather than those building its infrastructure.28

References

Every source behind this edition. Open one to read the full story.

  1. 1Thinking Machines wows with Inkling 2 Mini looped transformer model that runs locally on laptopsVentureBeat · venturebeat.com
  2. 2Introducing Inkling 2 Minithinkingmachines.ai · thinkingmachines.ai
  3. 3GPT-Live-1.1 has arrived in the APIcommunity.openai.com · community.openai.com
  4. 4The AI boom had a really bad week — and it's raising loud alarmsNPR · npr.org
  5. 5Meet Thinking Machines Inkling 2 Mini: 26B Looped MoE Runs on a 32GB Macmarktechpost.com · marktechpost.com
  6. 6thinkingmachines/Inkling-2-Mini · Hugging FaceHugging Face · huggingface.co
  7. 7OpenAI Ships GPT-Live-1.1 for Voice Agents, Cuts Idle Cost to $0.02/Minmarktechpost.com · marktechpost.com
  8. 8New GPT-6.2 Astra Builds a Working 3D Pendulum Clock from a Sketchgeeky-gadgets.com · geeky-gadgets.com
  9. 9Introducing GPT-6.2 Astra and SolOpenAI · openai.com
  10. 10China’s data regulator plans standards for measuring and billing AI tokensThe Next Web · thenextweb.com
  11. 11China plans new standards for AI token pricing, qualitytechinasia.com · techinasia.com
  12. 12Nvidia's neocloud empire lost its biggest champion. Washington may have dealt a bigger blow.Business Insider · businessinsider.com
  13. 13OpenAI says its agents now finish nearly half of 2 to 4-hour research tasksThe Next Web · thenextweb.com
  14. 14OpenAI Research | PublicationOpenAI · openai.com
  15. 15A tale of two NvidiasTechCrunch · techcrunch.com
  16. 16Google told schools to accept AI or lose Search. Now 3 of America’s 4 largest districts are pushing backFortune · fortune.com
  17. 17Anthropic’s record IPO collides with tech’s worst week on Wall Street since 2020CNBC · cnbc.com
  18. 18Google Is Paying Elon Musk's SpaceX $30 Billion. Jim Cramer Just Revealed Why247wallst.com · 247wallst.com
  19. 19One gigabyte of Wikipedia losslessly compressed to 105.39MB with self-learning AI transformer modelTom's Hardware · tomshardware.com
  20. 20tufazip: Learning to Compress a Billion Bytes of Wikipediatufalabs.ai · tufalabs.ai
  21. 21SC26: Phison Unveils Three Next-Generation Storage Innovations for the New Compute Era – PHISON Electronics Corp.phison.com · phison.com
  22. 22Phison shows an SSD design with an NPU and 64GB of memory that runs AI on the data it storesTechSpot · techspot.com
  23. 23Inside Early iOS 28 Leaks: Agent-Style Siri AI, Codename "Bell," and 2027 Hardware Supportgeeky-gadgets.com · geeky-gadgets.com
  24. 24A Team of Researchers from Columbia and CertiK Introduce TensorLean for Formal Verification of GPU Kernelsmarktechpost.com · marktechpost.com
  25. 25Nearly all software teams admit to shipping AI-generated code as study finds verification 'has not kept pace with that shift'flip.it · flip.it
  26. 26State of AI in Quality Engineering 2026: 81% merged AI code they didn't fully understandtestmuai.com · testmuai.com
  27. 27China’s tech giants double AI spending to $139bn as labs earn less than 8%The Next Web · thenextweb.com
  28. 281. FROM MY DESKtrk.wsj.com · trk.wsj.com

That’s the edition.

The next one is ready tomorrow by 6 AM ET.

Written with AI from the sources in the references above.

Ramit Koul
AuthorRamit KoulFounder, Software Engineer & Innovator