Good Morning!
Yesterday's text traffic on OpenRouter fell 15.1% to 24.91 trillion tokens, and almost every model on the board slipped with it.
One moved the other way: Kimi K3.1 Pro climbed from 12th to 5th, up 56.4%, after Moonshot published the weights.
DeepSeek V4.1 Flash still leads at 9.11 trillion, more than the next three combined, with 74.3% of its traffic machine-generated.
Model | tokens · vs. previous day (Nov 6) | ||
1 | DeepSeek V4.1 Flash · DeepSeek | 9.11T | – -12% |
2 | Hy4 · Tencent | 3.01T | – -1% |
3 | GLM 5.3 Flash · Z.ai | 1.28T | – -25% |
4 | GPT-6 Luna · OpenAI | 0.85T | – -27% |
5 | Kimi K3.1 Pro · MoonshotAI | 0.82T | ▲7 +56% |
6 | Claude Opus 5.5 · Anthropic | 0.77T | ▼1 -27% |
7 | crouton-alpha · stealth | 0.71T | ▼1 -32% |
8 | DeepSeek V4 Flash · DeepSeek | 0.64T | ▼1 -12% |
9 | MiniMax M3.1 Flash · MiniMax | 0.59T | ▼1 -14% |
10 | GPT-6.1 Sol · OpenAI | 0.56T | ▼1 -17% |
Ling 3.2 Flash Goes Multimodal

TL;DR: Ant Group launched Ling-3.2-flash with native text, image and video input, 560 billion total parameters, 25 billion active parameters and a one-million-token context window. It targets visual reasoning and long-running multimodal agents. Chat and API access are live, with OpenAI- and Anthropic-compatible interfaces and free use for the first two weeks after launch. Read more →
DSpine Speeds Up Qwen

TL;DR: Tencent released DSpine draft weights for Qwen3.8-27B, generating 16-token candidate blocks in parallel for the main model to verify. In the team's eight-request concurrent Office replay test at low reasoning effort, output throughput reached 3.84 times ordinary autoregressive decoding. The five-layer draft model and inference code are available under Apache 2.0, with SGLang deployment instructions. Read more →
Kanana Adds Visual Decisions

TL;DR: Kakao released Kanana-2-v-decide-3b-2611, a 3.67-billion-parameter model that returns probabilities for supplied choices, yes/no questions or ordered ratings. It accepts text, JSON or chat history with up to ten images and 20,480 tokens, supporting tasks such as routing customer requests and assessing attached photos. Weights are downloadable under the Kanana Open License. Read more →
Nous Retrains Qwen's Political Responses

TL;DR: Nous Research released Arete-27B, a Qwen3.8-27B fine-tune. In Nous's 803-prompt political-response evaluation, answers classified as propaganda fell from 22% to 0.9% after supervised fine-tuning on 52.8 million tokens over two epochs. Hermes Bench stayed nearly flat at 65.60% versus 65.77%. The weights are available under Apache 2.0. Read more →
ByteDance Opens Cola2 Weights

TL;DR: ByteDance Seed and HKU published Cola2 weights and inference code, pairing continuous semantic compression with a diffusion-based latent generator and a text decoder. The released base checkpoint was trained on 346.03 billion source tokens and averaged 60.10 across eight tasks in the team's evaluation. It is a research model rather than an instruction-tuned chat assistant. Read more →
SageMaker Adds GPU Partitioning

TL;DR: Amazon SageMaker Inference now supports NVIDIA Multi-Instance GPU for inference components, splitting supported accelerators into up to seven hardware-isolated partitions. Several models can share one GPU, or a small model can use only part of a newer accelerator. Partition layouts are fixed when an endpoint is created; AWS's testing found roughly 17 seconds of additional provisioning time. Read more →
Skymizer Plans Single-Server Kimi

TL;DR: Skymizer and MiTAC plan to demonstrate the 2.8-trillion-parameter Kimi K3 on one air-cooled server at SC26. Eight HTX301 LPU cards provide 2 TB of memory, designed to accommodate roughly 1.56 TB of weights plus KV cache; final memory allocations and supported workloads remain under engineering validation. HTX301 production is expected to begin in Q1 2027. Read more →
Chinese Models Lead Open Model Spend

TL;DR: Menlo Ventures' survey of nearly 400 US enterprise AI leaders estimates Chinese models account for 89% of open-model spending. Open models represent just 4% of total LLM spending in the report. Average open-model token prices fell from $0.45 to $0.13 per million while usage grew 22.5-fold. These figures track model usage, including deployments through third-party inference providers. Read more →
Open Weights Close Cyber Gap

TL;DR: Vals AI's CyberOffense results, updated October 23, put DeepSeek V4.1 Flash 1017 at roughly 73% objective completion across two test networks, versus 81% for restricted GPT-6 Sol Daybreak. On the Nova network, DeepSeek cost under $3 per run versus roughly $22. Each run had a maximum cumulative budget of 500 million tokens, including repeated context. 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.