← All updates

Thinking Machines Explains 4-Bit Training

Plus: Runway Hosts Seedance Flash, Gemini 4 Carbon Tops MathArena

Good Morning!

OpenRouter text models routed 25.5T tokens in the week ending Nov 14, up 2.5% week over week. DeepSeek V4.1 Flash held first on 10.1T, up 11%; Tencent's Hy4 slid 29.7% to 2.1T; GPT-6 Luna gained 20.5%.

DIT
Model Ranking
SATURDAY
Top 10 text models by real-world token usage · Nov 14, 2026 (UTC)
Real-world use:
Agent
Code
General
Data
Model
tokens
WoW (11/7)
1
DeepSeek V4.1 Flash
· DeepSeek
10.11T
–
+11%
2
Hy4
· Tencent
2.11T
–
-30%
3
GLM 5.3 Flash
· Z.ai
1.33T
–
+4%
4
GPT-6 Luna
· OpenAI
1.03T
–
+20%
5
Claude Opus 5.5
· Anthropic
0.75T
▲1
-2%
6
DeepSeek V4 Flash
· DeepSeek
0.66T
▲2
+3%
7
MiniMax M3.1 Flash
· MiniMax
0.54T
▲2
-9%
8
MiMo-V2.6-Flash
· Xiaomi
0.50T
▲8
+61%
9
Nemotron 3.5 Super VL
· NVIDIA
0.50T
▲3
+14%
10
ernie-5.5-flash-preview
· baidu
0.50T
NEW
How we rank: tokens processed over the full 24 hours of Nov 14, 2026 (00:00–24:00 UTC), counted and classified by DIT Radar's internal model.
Excluded: StepFun Step 5 Preview, whose traffic showed an anomalous, non-organic pattern.
Source: OpenRouter. The ranking reflects DIT Radar's interpretation of the data.

Thinking Machines Explains 4-Bit Training

Mira Murati, CEO of Thinking Machines Lab
Photo: Philip Pacheco / Getty Images

TL;DR: Thinking Machines Lab detailed how it used native 4-bit forward passes to train Inkling 2 Small, from the start of pretraining through the last reinforcement-learning rollout. In its largest training jobs, the lab reports an 8% pretraining speedup and a 20% reinforcement-learning speedup compared with full MXFP8 training. Read more →


Runway Hosts Seedance Flash

A generated character in Runway’s Seedance 2.5 Flash launch video
Frame: Runway

TL;DR: Runway added ByteDance’s Seedance 2.5 Flash to its video platform. Runway says the Flash version generates clips faster and at a lower cost than the full Seedance 2.5. The announcement gives no credit rate for Runway users. Read more →


SGLang Lands on Trainium

An AWS Trainium3 chip on a circuit board
Image: Amazon

TL;DR: SGLang-Neuron is now in public preview on AWS Trainium2 and Trainium3. The open-source plugin supports prefix caching, a hierarchical KV cache and running prefill and decoding on separate nodes. The team demonstrated Kimi K3 across three Trainium3 nodes and reported 72% to 75% higher output throughput with its host-memory cache on GPT-OSS-120B at 32K context. Read more →


Memorilla Adds a Memory Cache

Illustration of a gorilla reading, from the Memorilla project
Image: Stanford SNAP

TL;DR: Stanford’s Memorilla now has an optional cache for the module that turns a document collection into 16 query-specific memory tokens for a frozen LLM. In the authors’ B200 test on one million documents, the cache cut per-query latency for the memory module from 74 ms to 8.1 ms. The code and Qwen3-8B checkpoints are available under Apache 2.0. Read more →


Intel Reports 5.9× on Finance Tests

Intel’s Crescent Island GPU card
Image: Intel via Geeknetic

TL;DR: Intel presented results from four double-precision financial simulations at its AI Summit in Seoul. The preproduction Crescent Island GPU delivered 1.92× to 5.9× the throughput of the Nvidia comparison card, according to reporting from the event. The workloads were separate from Intel’s AI inference tests. Read more →


Gemini App Adds Hydrogen for Free Users

Google’s Gemini 4 Hydrogen launch artwork
Image: Google

TL;DR: Google is bringing Gemini 4 Hydrogen to the Gemini app. Free users can get Hydrogen through automatic model selection when Google determines it is the right fit for a prompt. This gives free users a way to use the model in the consumer app alongside the AI Studio and API access covered earlier. Read more →


Gemini 4 Carbon Tops MathArena

MathArena logo
Image: MathArena

TL;DR: Gemini 4 Carbon took first place on MathArena’s overall leaderboard with 85.4% expected performance. Claude Fable 5.2 scored 82.8% and GPT-6.1 Sol scored 80.6%. MathArena estimates Carbon costs $15.70 per problem versus $1.07 for Sol, roughly 15 times as much. Read more →


llama.cpp 1.0 Adds Kimi K3.1 Pro Support

llama.cpp v1.0.0 release artwork
Image: ggml-org

TL;DR: llama.cpp and its underlying tensor library ggml both reached version 1.0. The release adds support for Kimi K3.1 Pro and Kimi K3.1 vision, plus a speech synthesis endpoint at /v1/audio/speech. The projects will use major version bumps to signal breaking API or ABI changes. Read more →


Top AI models. Lower costs.

Access GPT, Claude, Gemini, and more through one OpenAI-compatible API—and save up to 70% on AI API costs. Create your DIT account, get your API key, and start building.