Liquid AI LFM2-24B-A2B is built on liquid neural networks rather than a standard transformer stack, and the hosted 24B/2B-active MoE is priced at $0.03/$0.12 per Mtok on Together AI. That is cheaper than almost every 8B model in the registry, from a genuinely different architecture. OpenRouter carries the same $0.03/$0.12 pricing via the kilo listing. The hard limit is context: 32,768 tokens with 32,768 max output on Together AI. Inside Atlas, LFM2-24B-A2B belongs in the `small_model` slot for titles and summaries, not in the build agent.
What is a liquid neural network model and how does LFM2-24B-A2B differ?
Liquid AI's LFM2 is built on liquid neural networks rather than a standard transformer stack. LFM2-24B-A2B is the hosted 24B/2B-active mixture of experts from that line, priced at $0.03/$0.12 per Mtok on Together AI, which is a genuinely different architecture at a genuinely low price.
Liquid AI's bet is architectural, not just an efficiency tweak on a transformer. LFM2-24B-A2B runs 24B total with only 2B active per token, which is why a model of this class runs at edge latency and this price. Liquid's architecture is designed for on-device and memory-constrained deployment, so the hosted API and a local deployment behave more alike than they would for a transformer MoE. That property matters if your longer-term plan is to move inference on-prem: what you measure on Together AI at $0.03/$0.12 per Mtok is closer to what you will get locally than the equivalent comparison for a transformer would be.
How cheap is Liquid AI LFM2-24B-A2B compared to other small models?
Liquid AI LFM2-24B-A2B costs $0.03 input per Mtok, which puts it at the very bottom of the registry, undercutting DeepInfra's GPT OSS 20B ($0.03/$0.14) on output at $0.12 per Mtok. Together AI and the OpenRouter kilo listing both quote $0.03/$0.12.
LFM2-24B-A2B is priced below almost every 8B model in the registry despite carrying 24B total parameters, because only 2B activate per token. Against the nearest comparable, DeepInfra's GPT OSS 20B at $0.03/$0.14 per Mtok, LFM2-24B-A2B matches on input and beats it on output. Inside Atlas, that makes it a candidate for the `small_model` slot: set `"small_model": "togetherai/LiquidAI/LFM2-24B-A2B"` in atlas.json for titles and summaries at $0.03 per Mtok. Atlas fans out work to subagents that can run in the foreground or in parallel background sessions, and cheap subagent tokens are what make that fan-out something you actually use rather than something you ration.
Is a 32,768 token context enough for a coding agent in 2026?
Liquid AI LFM2-24B-A2B holds 32,768 tokens of context with 32,768 max output on Together AI. That is small by 2026 standards. LFM2-24B-A2B cannot hold a large file plus its dependencies, let alone a repo slice, which is the constraint that decides where it fits.
The 32,768 token window is the single most limiting fact about LFM2-24B-A2B. Atlas searches code with hybrid semantic and keyword retrieval fused by reciprocal rank fusion and indexes code by AST declarations using tree-sitter, not blind line windows, which means it retrieves well, but good retrieval into a 32,768 token budget still cannot represent a large module and everything it imports. The practical rule is explicit: keep the 32,768 token limit in mind and do not route Atlas's codebase-index summarization through this model on a large repo. Where the window is fine is short-horizon work, session titles, individual file summaries, and commit messages, where 32,768 tokens is generous.
Can Liquid AI LFM2-24B-A2B be your Atlas build model?
No. Liquid AI LFM2-24B-A2B activates only 2B parameters per token, which means it is a fast responder, not a reasoner. The registry marks LFM2-24B-A2B non-reasoning, and it should not be your build model. Keep a real coding model in the `"model"` slot in atlas.json.
2B active parameters and no reasoning mode is a hard combination to build on. LFM2-24B-A2B answers quickly and cheaply at $0.03/$0.12 per Mtok, and that is the trade it makes. Writing a correct multi-file diff requires holding invariants and reasoning about consequences, and the registry's non-reasoning classification for LFM2-24B-A2B is an accurate description of what it will not do. Atlas keeps that decision reversible: Atlas lets you switch the active model and provider on the fly with favorites and recents, and Atlas computes a unified diff for every file edit and surfaces it for approval before writing, so even a bad experiment is a rejected diff rather than a damaged repo.
How do you configure Liquid AI LFM2-24B-A2B in Atlas?
Configuring Liquid AI LFM2-24B-A2B in Atlas takes three steps: export TOGETHER_API_KEY, run `atlas models togetherai`, and set the `small_model` key. LFM2-24B-A2B resolves as the `LiquidAI/LFM2-24B-A2B` row at $0.03/$0.12 per Mtok.
Together AI is the primary route for LFM2-24B-A2B. Export TOGETHER_API_KEY, or run `atlas login` and pick Together AI, then run `atlas models togetherai` and find the `LiquidAI/LFM2-24B-A2B` row before editing atlas.json so a bad model id fails loudly. Set `"small_model": "togetherai/LiquidAI/LFM2-24B-A2B"` for titles and summaries at $0.03 per Mtok. OpenRouter carries the same model at the same $0.03/$0.12 per Mtok through its kilo listing if you are already routing through OpenRouter. Leave the `"model"` slot on something that can reason.
Setup
- 01Export TOGETHER_API_KEY (or run `atlas login` and pick Together AI).
- 02Run `atlas models togetherai` and find the `LiquidAI/LFM2-24B-A2B` row.
- 03Set `"small_model": "togetherai/LiquidAI/LFM2-24B-A2B"` in atlas.json for titles and summaries at $0.03 per Mtok.
- 04Keep the 32,768 token limit in mind: do not route Atlas's codebase-index summarization through this model on a large repo.
- 05Leave the `"model"` slot on a reasoning-capable model, since the registry marks LFM2-24B-A2B non-reasoning.
Frequently asked questions
- how much does liquid ai lfm2-24b-a2b cost
- Liquid AI LFM2-24B-A2B is $0.03/$0.12 per Mtok on Together AI, and the OpenRouter kilo listing carries the same $0.03/$0.12 pricing. That $0.03 input price puts it at the very bottom of the registry.
- what is the context window of lfm2-24b-a2b
- Liquid AI LFM2-24B-A2B holds 32,768 tokens of context with 32,768 max output on Together AI. That is small by 2026 standards and cannot hold a large file plus its dependencies.
- what is a liquid neural network model
- Liquid AI's LFM2 is built on liquid neural networks rather than a standard transformer stack. The hosted LFM2-24B-A2B is a 24B/2B-active MoE, and the architecture is designed for on-device and memory-constrained deployment.
- is lfm2-24b-a2b a reasoning model
- No. The registry marks LFM2-24B-A2B non-reasoning. With 2B active parameters per token it is a fast responder, not a reasoner, and it should not be your build model.
- how do I use lfm2-24b-a2b with atlas
- Export TOGETHER_API_KEY, run `atlas models togetherai` and find the `LiquidAI/LFM2-24B-A2B` row, then set `"small_model": "togetherai/LiquidAI/LFM2-24B-A2B"` in atlas.json for titles and summaries at $0.03 per Mtok.
- lfm2-24b-a2b vs gpt oss 20b pricing
- LFM2-24B-A2B is $0.03/$0.12 per Mtok on Together AI. DeepInfra's GPT OSS 20B is $0.03/$0.14 per Mtok. They match on input, and LFM2-24B-A2B undercuts GPT OSS 20B on output.
- cheapest model for atlas titles and summaries
- Liquid AI LFM2-24B-A2B at $0.03/$0.12 per Mtok is one of the cheapest. Its 32,768 token window is enough for titles and summaries, but do not route Atlas's codebase-index summarization through it on a large repo.
Try Atlas in your terminal
The terminal-native AI coding agent. Free core, single binary.
Install AtlasRelated guides
Audit a Repo with Parallel Subagents in Atlas (2026 Workflow)
How to audit a repo with parallel subagents in Atlas in 2026: the task tool launches explore subagents in their own sessions, so only conclusions return to your context.
Atlas for Perl: A Terminal-Native AI Coding Agent for CPAN Distributions in 2026
Atlas is a terminal-native AI coding agent for Perl in 2026. It reads cpanfile deps and @EXPORT lists, writes Test2::V0 cases, runs prove -lr t/, and runs perltidy on the diff.
Atlas vs Greptile: Terminal AI Coding Agents in 2026
Comparing Atlas and Greptile in 2026. Atlas offers terminal-native AI coding with permission-gated tools. Greptile reviews code with sandbox execution, catching 20% more bugs.
Atlas for Astro: Islands, Content Collections, and Zero JS by Default in 2026
Atlas is a terminal-native AI coding agent for Astro in 2026. It reads astro.config.mjs, src/pages, and content collection schemas, drops needless client:load directives, and runs astro check.
Atlas vs Traycer in 2026: Terminal Agent That Writes Code vs a Planning Layer Above One
Atlas vs Traycer in 2026. Traycer plans and verifies but writes no code itself, so you pay it on top of an agent. Atlas plans and writes, free and open source.
Atlas vs Qwen Code: Terminal AI Coding Agents in 2026
Atlas and Qwen Code are leading terminal AI coding agents in 2026. Compare Atlas's TUI, permission-gated tools, and free core with Qwen Code's 1M token context, Agent Teams, and Alibaba Cloud hosting.
Atlas for TypeScript in 2026
In 2026, TypeScript developers leverage Atlas, the terminal-native AI coding agent, to enhance productivity. Atlas understands your types, ensures code quality, and offers robust safety features.
Atlas for Haskell in 2026
Atlas is a terminal-native AI coding agent for Haskell in 2026. Run it in a project with a .cabal file or a stack.yaml, let cabal build drive it, and review the diff.