Theme

The frontier gap becomes a measured latency

Explore this theme’s evidence over time →

Sources

Attribution and reporting linked by this record; links do not imply independent verification.

Welcome to December 23, 2025 (theinnermostloop.substack.com)

Open-weight models trail the closed frontier by a number of months that can be stated and tracked. The moat is re-described as a release schedule.

Development timeline

Newest reporting dates first. Reporting dates may differ from event dates.

2026-08-15: An open model nearly matches the frontier at vulnerability discovery

2026-08-06: A weather model buys an extra day of cyclone lead time and is open-sourced

2026-07-18: Top open models trail the cyber frontier by just four to seven months

2026-07-17: An open model matches the frontier on kernels and autonomously designs a chip

2026-07-17: A researcher says a 300-person lab cracked the compute-moat thesis

2026-07-16: The strongest open weights in the West are dethroned within a day

2026-07-12: A single-file engine hosts a 744B model on a consumer laptop

2026-07-12: A researcher concludes there is no moat, only shared recipes and migration

2026-07-09: Near-frontier agentic coding arrives from an open base at 1000 tokens a second

2026-07-07: One model leads all eight industry indices at over 100x the cost

2026-07-04: Cheap post-training makes sovereign models viable for everyone

2026-07-01: Local models are found able to handle 88.7% of chat queries

2026-06-29: Companies rush to post-train and own their own weights

2026-06-28: An MIT-licensed model folds 1.6 trillion parameters into a tenth the KV cache

2026-06-27: Analysis projects the open-versus-closed gap closing to zero

2026-06-25: A founder open-hands a humanoid platform as too vital to gatekeep

2026-06-24: An open model fixes a real bug at half the cost and more tidily

2026-06-22: A sovereign initiative ships a fully open model in 1,000+ languages

2026-06-20: Chinese open models take the majority of token use

2026-06-19: A 744B open model is shrunk from 1.51TB to 217GB to run locally

2026-06-17: A Chinese open model takes the lead while the frontier sits unavailable

2026-06-15: An open model beats the frontier on ML engineering under cost limits

2026-06-14: A municipal IT office post-trains a state-of-the-art coder

2026-06-14: A Chinese lab answers the shutdown with radical openness

2026-06-12: Nations are offered a way to co-train one shared model with local data

2026-06-04: An encoder-free multimodal model runs on a laptop under Apache 2.0

2026-06-02: An open-weight model undercuts the frontier forty-fold

2026-06-01: A compact image model runs on a phone

2026-05-18: Chinese labs overtake the US in video generation

2026-05-15: The first mixture-of-experts model is trained end to end off Nvidia

2026-05-12: The White House weighs banning Chinese cellular modules

2026-05-11: A rival software stack improves 75-fold in fourteen days

2026-05-01: Google concedes it lacks the silicon for two frontier model families

2026-04-26: An open 1.6T model puts China four to five months behind the frontier

2026-04-26: The Free Software Foundation calls responsible AI licences unethical

2026-04-09: A model crushes data-quality benchmarks and flubs reasoning ones

2026-04-08: Meta Superintelligence Labs ships its first model

2026-04-03: Small open models outcompete rivals twenty times their size

2026-04-02: The near-term framing shifts from superintelligence to replication

2026-04-01: A single-bit 8B model matches full precision in 1.15 GB

2026-03-31: Alibaba's newest omni-modal model ships closed

2026-03-24: A 400-billion-parameter model runs on a phone

2026-03-15: An open agent scopes, derives, verifies and writes up physics research

2026-03-15: An autopilot mode compresses a well-scoped problem from weeks to hours

2026-03-12: Nvidia sees no wall in post-training and opens a 120B model

2026-03-03: Four-billion-parameter models match last generation's eighty

2026-03-02: A full AI stack boots in 8 ms on a $5 microcontroller

2026-02-13: Agentic capability arrives at a dollar an hour

2026-02-12: Agents meta-learn their own memory architecture

2026-02-05: An ensemble of three labs' models hits 94.5% on ARC-AGI

2026-01-27: Chinese models match the frontier on nineteen benchmarks

2026-01-24: A 10B Chinese model claims to beat models twenty times its size

2026-01-14: The US allows H200 sales as China forbids buying them

2026-01-07: An 8B model reaches SOTA forecasting by predicting its own future

2026-01-06: Nvidia puts a fifth of its workforce on autonomous driving

2026-01-04: ByteDance takes the top spot on FoldBench

2026-01-01: A 40B looped transformer claims 81.4% on SWE-bench Verified

2025-12-31: Decentralized training compute grows four times faster than frontier runs

2025-12-29: GLM-4.7 takes the top open-weight slot

2025-12-29: A 32B Korean model beats Gemini 3 Pro on tool use

2025-12-28: GLM 4.7 becomes the first open-weight model to profit on Vending-Bench 2

2025-12-24: MiniMax M2.1 targets Rust, Java and C++

2025-12-23: Epoch puts the open-weight FrontierMath lag at seven months

2025-12-23: GLM-4.7 tops GPT-5.2 on WebDev Arena, trailing the frontier by six months

Knowledge graphopen in browser →

Cite or correct this record

Citation formatCitation text

The frontier gap becomes a measured latency
Innermost Loop KB — secondary record
Reporting date: unknown
KB permalink: https://nicholsn.github.io/innermost-loop-kb/themes/open-weight-latency
Attributed sources:
- Welcome to December 23, 2025 — https://theinnermostloop.substack.com/p/welcome-to-december-23-2025
This citation does not imply independent verification of linked-source claims.

Suggest a correction — opens a GitHub draft for you to review and submit.