Free Intensive Nov 21–22Free 2-day hands-on Vibecoding Intensive on Nov 21–22 at 16:00 MSK (Sat–Sun) · Live Cohort 3 starts Nov 23Join Free →Cohort 3 (Nov 23)
ReleaseAPI / Product · ActionableХабр: NLP→ original

Jev Returns Decisions 40–200 Times Faster Than LLMs, Google Answers with a Counterpart

Agents often need not text but a fast choice: which tool to call, or whether an action is risky. TypeSafe AI has released Jev, a model that takes a description of a situation and a list of questions and returns answers with probabilities in a single pass, without parsing text. According to the company, it is 40–200 times faster than conventional LLMs on such tasks. In November, Google responded with Gemini Decision 1, so models like these are worth a look for fast decisions in agents.

AI-processed from Хабр: NLP; edited by Hamidun News
Jev Returns Decisions 40–200 Times Faster Than LLMs, Google Answers with a Counterpart
Source: Хабр: NLP. Collage: Hamidun News.
Listen to article
Article Fact Sheet & Key Figures
ParameterDetails
Primary sourceХабр: NLP ↗
Editor-in-Chief
Gemini Decision 1150–250 ms latency, $0.10 per million input tokens
Key metric #2Up to 1000 questions per request and a 256K-token context window in Google's model
Key metric #3Jev's speed is the vendor's claim on automation tasks, not on all tasks

Picture a support agent who receives a customer email. It needs to decide three things: refund the money, hand the request to a human, or simply reply, and whether an additional check is required. A conventional LLM would write a paragraph of text for this, which then has to be parsed.

In September 2026, TypeSafe AI proposed a different approach: its Jev model does not generate text at all. It receives a description of the situation and a list of questions with predefined answer options, and immediately returns a choice along with a probability. According to the company itself, on suitable automation tasks this is 40–200 times faster than traditional LLMs.

Then on November 11, Google entered this niche with Gemini Decision 1. Let's look at what these models are and which other alternatives to GPT have already come out of the labs.

Models that don't write text: Jev and Gemini Decision 1

GPT-like LLMs have two defining traits: a decoder-only Transformer architecture and autoregressive generation, in which each next token is predicted from the previous ones (like T9, but far more powerful). These are independent properties, and either one can be replaced.

Jev drops sequential generation wherever an application needs a specific answer: choosing a tool, classifying a request, assessing risk, determining the next step. Questions are processed in parallel, and the result comes back structured, together with probabilities. The model is trained with the Reinforcement Learning for Calibrated Decisions (RLCD) method. It learns not only to pick the correct option but also to honestly estimate its confidence in that option: if the model says it is "90% sure," it should be right about 90% of the time.

The figure of 40–200 times should be read carefully. It is a vendor's comparison on specific scenarios, not a universal AI speedup. The company has not yet disclosed Jev's internal architecture, so it is too early to classify it as an SSM or a linear-attention model. Jev also does not aim to replace general-purpose LLMs: one model may write code, while another quickly decides which tool to call.

Gemini Decision 1 works on the same principle: in a single pass it evaluates options and returns probabilities and confidence scores. But it is multimodal: it understands text, images, audio, video, and PDFs, handles a context of up to 256 thousand tokens, and accepts up to 1000 independent questions per request. Google claims a latency of 150–250 ms and a price of $0.10 per million input tokens, with output tokens not billed. According to the company's own tests, the model outperforms Jev on some benchmarks, but not on all of them. Google uses it to route requests between models and to check the safety of tools.

Recurrent networks, linear attention, and hybrids

Recurrent networks have not disappeared; they have been rebuilt. RWKV keeps a recurrent state during inference while training in parallel. Despite the Key and Value in its name, there is no growing KV-cache (memory of past keys and values) in it. xLSTM updates the classic LSTM with exponential gating and matrix memory. The NXAI team trained a 7-billion-parameter model and released the code and weights. But open weights do not mean the model will easily fit into your inference server: it needs specialized kernels, quantization, and proper batching, and here the Transformer has a major advantage.

Linear attention compresses history into a compact state so that the keys and values of every token do not have to be stored. The main question is how not to erase what matters. DeltaNet checks the association already stored before writing and corrects the memory. Gated DeltaNet controls forgetting, and Kimi Delta Attention (KDA) does this per channel. In 2026, NVIDIA proposed Gated DeltaNet-2, where erasure and writing are regulated separately.

In practice, hybrids are more common. Qwen3-Next-80B-A3B contains 80 billion parameters, of which about 3 billion are active per token, and its Gated DeltaNet layers alternate with global attention. The gap between total and active parameters is due to MoE, not to a new architecture. In Moonshot AI's Kimi Linear 48B-A3B, one full-attention layer follows every three KDA layers: the cheap blocks process most of the sequence, while the global layers keep direct access to the context. These models still keep a KV-cache for their global layers, so the real savings depend on batch size, text length, and the quality of the compute kernels.

Convolutions and diffusion: LFM3, LLaDA, Mercury

Liquid AI has shown that convolutions also work for text. In LFM2, short convolutions are combined with separate attention layers, and the architecture was tuned for real hardware, including mobile CPUs. On October 27, 2026, LFM3 was released for text, vision, and audio. The text model, LFM3-1.7B, has 24 layers: 13 convolutional, 5 with local attention, and 6 with global attention. This is not a model without attention; attention is simply not needed in every layer. The goal is local agents operating under tight memory, energy, and latency limits.

Diffusion LLMs change the order of generation. The model starts from a partially masked sequence and refines many positions at once over several steps. LLaDA has shown competitive results on a number of tasks at the 6–10B scale, although it still uses a Transformer internally; the difference lies in non-autoregressive generation. Inception Labs' Mercury has reached a commercial API and code. Parallel refinement delivers high throughput in the published tests, but the extra iterations cost compute, so the gain depends on response length and hardware. For code, the idea is appealing: edits are often needed in the middle of a file. However, the quality of such editing has to be verified on real diffs.

What developers should do

Do not confuse MoE, quantization, speculative decoding, and RAG with new architectures: these are ways to make LLMs cheaper, not replacements for GPT. Choose based on the product's constraint. Streaming workloads point to recurrent models, long repositories to hybrids, local assistants to compact models that run on CPUs or NPUs, and fast decisions inside an agent to models like Jev and Gemini Decision 1.

So far, neither RWKV, linear attention, nor diffusion has universally displaced the Transformer, which has mature infrastructure and proven scaling. A more useful question is not "which model is smarter" but "which architecture, on which hardware, with which agent scaffolding, best solves my task." All the speedup figures here are claimed by vendors, so verify them on your own workloads.

Story

Key Questions Answered

What are the key details on Models that don't write text: Jev and Gemini Decision 1?

GPT-like LLMs have two defining traits: a decoder-only Transformer architecture and autoregressive generation, in which each next token is predicted from the previous ones (like T9, but far more powerful). These are independent properties, and either one can be replaced. Jev drops sequential generation wherever an application needs a specific answer: choosing a tool, classifying a request, assessing risk, determining the next step.

What are the key details on Recurrent networks, linear attention, and hybrids?

Recurrent networks have not disappeared; they have been rebuilt. RWKV keeps a recurrent state during inference while training in parallel. Despite the Key and Value in its name, there is no growing KV-cache (memory of past keys and values) in it. xLSTM updates the classic LSTM with exponential gating and matrix memory. The NXAI team trained a 7-billion-parameter model and released the code and weights.

What are the key details on Convolutions and diffusion: LFM3, LLaDA, Mercury?

Liquid AI has shown that convolutions also work for text. In LFM2, short convolutions are combined with separate attention layers, and the architecture was tuned for real hardware, including mobile CPUs. On October 27, 2026, LFM3 was released for text, vision, and audio. The text model, LFM3-1.7B, has 24 layers: 13 convolutional, 5 with local attention, and 6 with global attention. This is not a model without attention; attention is simply not needed in every layer.

ZK
Zhemal Khamidun — CPO at Alpina Digital, product manager of AlpinaGPT. Building AI products since 2019. Hamidun News is a pet project that grew into a production aggregator with 6,000+ readers.
✦ Unified AI Workspace · Nomad

Test Frontier LLMs & Autonomous Agents in One Workspace

Run Claude Opus 5.5, GPT-6, Gemini 3.1, and DeepSeek side-by-side in Nomad AI (nomadnet.ai) — unified model routing, deep research, and multimodal workflows under one subscription.

Try Nomad AI →Launch Nomad Agent
What do you think?
Loading comments…