← All updates

Five Models Open Weights

Plus: Microsoft Opens Aion Weights, Mistral Vibe Free in EU,

Good Morning!

Tencent's Hy4 Flash took 44% of OpenRouter's text traffic last week. Today the weights are out: the best-selling model on the market is now free to download.


DIT
Model Ranking
WEDNESDAY
Top 10 text models by real-world token usage · Nov 18, 2026 (UTC)
Real-world use:
Agent
Knowledge
Chat
Model
tokens
WoW (11/11)
1
Hy4 flash
· Tencent
23.81T
NEW
2
DeepSeek V4.1 Flash
· DeepSeek
12.31T
▼1
+2%
3
GLM 5.3 Flash
· Z.ai
1.93T
–
-26%
4
GPT-6 Luna
· OpenAI
1.49T
–
+19%
5
Hy4
· Tencent
1.34T
▼3
-50%
6
Claude Opus 5.5
· Anthropic
1.14T
▼1
+10%
7
Kimi K3.1 Pro
· MoonshotAI
0.68T
▲4
+31%
8
DeepSeek V4 Flash
· DeepSeek
0.67T
▼2
-31%
9
MiMo-V2.6-Flash
· Xiaomi
0.61T
▼1
-9%
10
GLM 5.3
· Z.ai
0.61T
▼3
-24%
How we rank: tokens processed over the full 24 hours of Nov 18, 2026 (00:00–24:00 UTC) on OpenRouter. Usage lanes follow OpenRouter's task classification, grouped by DIT Radar into Agent / Knowledge / Chat and distributed to each model by our estimate.
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.

Hy4 Flash Open-Weights

Photo: Donald Wu, Unsplash

TL;DR: Tencent has released open weights for Hy4 flash, including Hy4 flash-FP8 and Hy4 flash-Base. The natively multimodal model totals 224B parameters with 12B active and a 1M-token context window, positioned for fast, efficient everyday work. Teams can now self-host or fine-tune a model that was API-only earlier this week. Read more →


Microsoft Opens Aion Weights

Microsoft campus in Redmond, Washington
Photo: David Ryder / Bloomberg via Getty Images

TL;DR: Microsoft released Aion 1.0 Instruct, a 2.3B-parameter model for local summarization, rewriting and chat, under the MIT license. It supports eight languages and a 32K-token context window. The weights are available now; Microsoft plans an optimized version with content moderation as Windows' built-in small language model by the end of January 2027. Read more →


Exa Embedding Speed +50%

Exa co-founders Jeff Wang and Will Bryk
Photo: Exa

TL;DR: Exa raised embedding generation throughput by 50 percent running on AWS AI chips, indexing billions of documents. Faster, cheaper embedding throughput directly lowers the cost of large-scale retrieval pipelines, a persistent line item for search and RAG products. Read more →


Cloudflare Frees Clairic

Cloudflare office entrance at 101 Townsend Street in San Francisco
Photo: Sundry Photography / Shutterstock

TL;DR: Cloudflare's new Workers AI superintelligence team shipped an updated Clairic and the new Clairic-flash as open-weight extraction models, also hosted on Workers AI. The 27B Clairic lists at $0.24 per million input tokens and the 9B flash at $0.10, with a $0.10 promotional rate until December 1, 2026; both bill input tokens only. Read more →


AMD GPU IP Gains 14% PPA

Photo: Ilias Gainutdinov, Unsplash

TL;DR: AMD says models built with OpenAI identified up to 14 percent power-performance-area improvement on its flagship GPU IP in under five months, building on years of in-house optimization. AI-assisted chip design is becoming a measurable lever in inference hardware roadmaps. Read more →


DaisyRose-27B Goes Open

Image: Hugging Face

TL;DR: RadixArk open-sourced DaisyRose-27B under Apache 2.0, its first model, trained from Qwen3.8-27B with reinforcement learning on long-horizon agent tasks. Terminal-Bench 4.1 pass@1 rises from 5.6 to 12.0 percent; the run used 1,200 curated tasks over 18 days on 144 H200 GPUs and cut training cost 57 percent versus its open-source baseline. Read more →


Rumble-2-Flash Goes Open

Image: Hugging Face

TL;DR: Rumble-2-Flash is a new open-source speech model under Apache 2.0: 0.5B parameters covering more than 40 languages. Small, permissively licensed voice models keep pushing on-device and low-latency speech use cases within reach of commodity hardware. Read more →


Mistral Vibe Free in EU

Photo: Pavel Neznanov, Unsplash

TL;DR: Mistral is making its Vibe model free in Europe for the first three months, crediting the Vibe plus Le Chonk harness pairing as the strongest open coding setup. A time-boxed free tier aims squarely at European developers weighing open alternatives to GLM and DeepSeek. 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.