Large Language Models

coding, long-running tasks and cybersecurity, with restricted initial access

visual reasoning, computer use and writing at $1 input and $5 output

a reasoning model for coding, 3D design and agent tasks

a diffusion language model with a million-token context

a low-cost model for coding agents

An open-weight update for coding and tool-using agents

Downloadable weights for coding and agents

downloadable weights for multimodal agents

vision-language models for agents

Liquid AI’s LFM2.5-24B-A2B is a reasoning model for local agents. Its hybrid architecture, 131,072-token context, downloadable weights and licence conditions.

Meta's models for coding and longer tasks

scoring choices for software

Celeris-1 Decision evaluates supplied choices and returns structured decisions with probabilities. Its hosted API launched on 8 October 2026.

JetBrains Mellum2.1 is a coding model released on 8 October 2026. It has 12B total parameters, 2.5B active parameters and Apache 2.0 weights.

Claude Haiku 5.5 is Anthropic’s small model for summaries, classification and routine coding work. Its prices depend on whether a prompt exceeds 100,000 tokens.

Liquid d1 models score possible answers to a question. The open d1-3B and experimental d1-omni-600M releases bring decision models to local text, image and audio workloads.

A trillion-parameter model opens in public preview

classifying text against supplied rules

One compact model puts different media into a shared search space

Kolibri-1 is Aleph Alpha’s downloadable language model for German and English

Ai2 released AstaBrief 8B on 2 October 2026

Cloudflare's first trained models: open decision models that answer typed questions with probabilities

the decision model behind Perplexity's Decisions API, with open weights

a small open decision model that runs on a gaming GPU or a laptop

one endpoint that runs several models on every request, now 57% to 88% cheaper per token

longer reasoning, coding and cyber defence

Bilibili’s Index-Translate models translate text across 150 languages and follow terminology and formatting instructions

Ling 3.1 Flash is InclusionAI’s language model for coding, analysis and tool use

Cohere's embedding models for search over text, images and documents

the upgraded mid-price GPT-6 model, released 29 September 2026

Nvidia’s Kumo Tabular predicts categories or numbers from rows of data

faster everyday coding and office work at Sonnet prices

Naive-N0.5-Flash is a downloadable language model with 309 billion parameters, 15.5 billion active per token and a one-million-token context. Its weights use the MIT licence.

Perceptron Mk1.5 reads text, images, video and audio for applications such as robots and smart glasses. It returns text, spatial annotations and tool calls through a hosted API.

Liquid AI’s LFM2.5-VL-DSpark helps its 3B vision-language model write answers faster

multi-model roleplay and storytelling systems, released 23 September 2026

the mid-price and budget GPT-6 models, released 22 September 2026

the first Claude 5.5 model, cheaper and faster than Opus 5

Upstage's small, cheap model for agents, released 22 September 2026

xAI's current flagship, and the cheapest model on its own comparison table

Flash, Pro and Pro UltraSpeed, released together

StepFun's flagship for long agent jobs, with downloadable weights

the live dialogue pair, one built for scale and one for thinking

a decision model that returns a typed answer and a confidence

a 552B open-weight model built around its cache

the new generation, rolling out from 3 September 2026

a six-model Apache 2.0 fleet released with its checkpoints, logs and datasets

Meta's proprietary frontier family

the workhorse again, plus a cyber variant behind a gate

built for the API, hard on a subscription allowance

an open-weight flagship on a standard open-source licence

the cheap tier, and a preview of what Qwen4 is built on

a new base, natively multimodal, at a tenth of GLM-5.2's price

the August experimental vision release

an anonymous listing, measured rather than guessed

a post-training update with an emergent cyber capability

August’s Flash release and its API successor

the price disruptor, raising its prices

xAI's flagship until Grok 4.7 replaced it, built for long-running agents

Downloadable models for local applications

the first Qwen Max released with open weights

frontier work with a dial on the bill

Google’s May 2026 generation

the efficient workhorse

the largest open-weight model yet released

the newest lab's first big swing

the flagship from 9 July to 3 September 2026

the coding pivot, superseded by Grok 4.6

the ecosystem model

the speed and intelligence balance

China's biggest consumer assistant

strongest open-weight all-rounder before K3 landed

the run-up to K3

the first model in Anthropic's Mythos class

restricted twin of Fable 5

the on-device giant

the transparency play

open-weight agentic climber

the benchmark everyone shoots at

Qwen's premium, API-only tier

Cohere's open-weights flagship for business agents

retrieval and RAG for enterprises

the incumbent's answer

the Gulf's open flagship

compliance-first open models

the open-weight price disruptor

recent predecessor, still widely deployed

the multilingual open workhorse

the laptop model

small and on-device

the superseded mid-tier workhorse

Europe's flagship

the Bedrock native

fully open science

the fast cheap tier

the SSM-transformer hybrid

the community's default base

search-native answering