View all articles
newsopinionno-hypeai

No Hype: Kimi K3.1, Hy4, ERNIE 5.5, Ling 3.2 and Ripple-65B

JG
Jacobo González Jaspe
|
No Hype: Kimi K3.1, Hy4, ERNIE 5.5, Ling 3.2 and Ripple-65B No Hype
This article is also available in Spanish:Sin humo: Kimi K3.1, Hy4, ERNIE 5.5, Ling 3.2 y Ripple-65B

Moonshot AI published the weights of Kimi K3.1 Pro, a 3T-class model with a 1-million-token window, and four more big models followed it this week. Here is each one, with the numbers its makers and one independent tester published.

Moonshot AI publishes the weights of Kimi K3.1 Pro

Illustration generated with AI on our own machine. Illustration generated with AI on our own machine.

What happened. Moonshot AI released Kimi K3.1 Pro, an open-weight, natively multimodal agentic model that it calls its most capable 3T-class model to date (Hugging Face). It has a 1-million-token context window. The model card’s footnotes show their working: the ExploitBench run used a 300-turn budget, and some rival scores were taken from DeepSeek’s own release blog.

What we think. Open weights at this size are good news for anyone who cares where their data goes, and bad news for anyone hoping to run them on a laptop. A 3T-class model is a server-room purchase. We do like footnotes that say where each comparison score came from.

The good bit. You can download it, which means you can host it, audit it and keep your documents on your own network.

What you can do. Put the size into our VRAM calculator before you get excited. If the answer is a rack, a smaller open model will read your invoices just fine.

Tencent opens the Hy4 preview, a 770B mixture-of-experts model

Illustration generated with AI on our own machine. Illustration generated with AI on our own machine.

What happened. Tencent’s Hy team released the Hy4 preview, a mixture-of-experts model with 770B total parameters, of which 49B are active per token (Hugging Face). Artificial Analysis tested it through Tencent’s API: 35 on its Intelligence Index against a median of 37 for its price tier, $0.83 per 1M input tokens and $2.50 per 1M output tokens, and a time to first answer token of 103.27 s against a median of 14.55 s (Artificial Analysis).

What we think. Cheap and slow is a real trade, and the independent test shows both halves of it. Tencent’s own blind test had 163 internal experts rate 203 engineering tasks, with Hy4 at 2.99 against 2.94 for Kimi K3. Staff grading their employer’s model is a press release with a spreadsheet attached.

The good bit. The weights are public, so you are not tied to Tencent’s API or its queue.

What you can do. For overnight batch work that nobody watches, the price matters more than the wait. For anything a customer waits on, it does not. Match the job to the machine with our hardware by job guide.

Baidu releases ERNIE 5.5 Flash Base for further training

Illustration generated with AI on our own machine. Illustration generated with AI on our own machine.

What happened. Baidu’s ERNIE team announced ERNIE-5.5-Flash-Base on 5 November 2026 (Unite.AI). It is a mixture-of-experts model with 280B total parameters and 14B active per token, pre-trained on 33.5T tokens, with a training context of up to 128K and 1M tokens supported for fine-tuning and inference (ERNIE).

What we think. A base model is raw material, not a product: it is meant for teams that will train it further. Baidu says the design roughly halves expert routing traffic and expert FLOPs, and its full training setup uses 16 nodes with four GB300 GPUs each (Unite.AI). That setup tells you exactly who the audience is, and it is not a dental clinic in Russafa.

The good bit. Baidu publishes the architecture in detail, such as layers that compress every 128 tokens into a single cached entry (ERNIE), so engineers can learn from it without running it.

What you can do. Unless you plan to fine-tune a 280B model, skip this one. If you build on local models, start from our local AI for developers page.

Ant Group adds Ling 3.2 Flash to its open Ling series

Illustration generated with AI on our own machine. Illustration generated with AI on our own machine.

What happened. Ant Group, which develops and open-sources the Ling series, released Ling-3.2-flash, the latest native multimodal model in the series, with 560B total parameters, 25B active, a context window of up to 1M tokens and image, text and video input (Ant Ling). The same page lists Ling-3.1-flash at around 560B total, about 25B active, also with up to 1M.

What we think. The size and the window are the same as 3.1 Flash. What changed is the input: 3.2 Flash reads images and video natively, which matters for scanned paperwork more than any parameter count.

The good bit. The series also goes small: Ling-3.0-tiny activates only 1.3B parameters (Ant Ling).

What you can do. For a small office, the tiny end of a series like this is the interesting end. For a proven local reasoning option, compare it with DeepSeek R1: The Best Open-Source Reasoning Model You Can Run Locally.

Ricursive Intelligence releases Ripple-65B Preview, trained for 32 million dollars

Illustration generated with AI on our own machine. Illustration generated with AI on our own machine.

What happened. Ricursive Intelligence released Ripple-65B Preview, a 65B dense Transformer pretrained on 12 trillion tokens with up to 1M tokens of context, under Apache 2.0 (Together AI). It is a general-purpose version of the in-house model behind Ricursive’s chip design platform. Ricursive says it trained the model for $32M and that it scored 80.52 on TerminalBench 2.1, ahead of NVIDIA Nemotron 3 Ultra 550B and Thinking Machines Inkling 975B (Ricursive).

What we think. This is the most practical model in the issue: 65B dense, and the model card reports 250 tok/s on 2xB200 before any quantization (Hugging Face). The team grew from 1 to 5 people. Sam Altman said in 2023 that GPT-4 cost more than $100 million to train (Ricursive), so $32M is a discount, if a very expensive one. The benchmark wins are Ricursive’s own numbers.

The good bit. Apache 2.0 lets you use it commercially, change it and run it wherever you like.

What you can do. Try it on Together AI at $0.20 per 1M input tokens and $0.60 per 1M output tokens (Together AI) before buying anything, then check what a 65B model needs in our VRAM calculator.

In one line

Kimi K3.1 Pro is open and enormous: download it if you have the servers for it, and pick a smaller open model if you do not.

Work with us

If something in this column belongs in your business, on models that run on your own hardware rather than someone else’s cloud, tell us what you need or see how we work in consulting.

Diagram
Share: LinkedIn X
Veredicto semanal

Get new guides before anyone else

Subscribe and we tell you when new guides, templates and workflows go up. One email a week, no spam.

Already published: 102 guides and 25 templates. All free, no signup.

Bonus: the local-AI starter pack PDF when you subscribe
Once a week No spam Unsubscribe anytime

See what you get