K-EXAONE 2.5, which LG AI Research opened to developers on November 9, 2026, is a free, OpenAI-compatible API for a 750-billion-parameter agentic model with 31 billion active parameters. It is text-only, supports Korean and English, always reasons before answering, and accepts up to 524,288 tokens per request.
Last updated: November 9, 2026
Table of contents
- What is new in K-EXAONE 2.5
- Specs compared with K-EXAONE 2.0
- How to make your first call
- Reasoning and sampling settings
- Using it with Hermes Agent and OpenCode
- Rate limits, errors and limitations
- Frequently asked questions
- Sources
What is new in K-EXAONE 2.5
LG AI Research announced on November 9, 2026 that the K-EXAONE 2.5 API is free to use through the EXAONE API Platform. LG describes the model as a technology-validation model built by applying additional training to the version it released in August. Anyone who obtains an API key can use it for service development and research without a separate fee.
The launch matters for power users for three reasons. It is a frontier-scale model with no per-token charge at launch, it speaks the familiar Chat Completions format, and LG published ready-made configurations for two popular agent harnesses on day one.
LG’s self-measured agentic benchmarks put K-EXAONE 2.5 close to open-weight models of similar scale. The models page lists these results:
| Benchmark (as reported by LG) | K-EXAONE 2.5 | K-EXAONE 2.0 | DeepSeek-V4.1-Flash | GLM-5.3 |
|---|---|---|---|---|
| SWE-Bench Pro (agentic coding) | 59.21 | 47.1 | 58.41 | 64.03 |
| BrowseComp (agentic search) | 78 | 50.87 | 75.8 | 79 |
| AutomationBench, open set (agentic tools) | 74.6 | 16.9 | 74.9 | 71.6 |
| Claw-Eval (agentic tools) | 77.29 | 62.11 | 75.78 | 77.64 |
For context on the comparison models, see our DeepSeek V4.1 Flash guide. Treat these as vendor-reported numbers rather than independent results.
Specs compared with K-EXAONE 2.0
The table below combines the EXAONE models page for version 2.5 with the K-EXAONE 2.0 model card on Hugging Face.
| Item | K-EXAONE 2.5 (API) | K-EXAONE 2.0 (open weights) |
|---|---|---|
| Parameters | 750B total, 31B active | 750B total, 37B active |
| Context length | 524,288 tokens (512K) | 262,144 tokens |
| Max output tokens | 512K | See the model card |
| Languages | Korean and English | Ten languages, including Korean, English, Spanish and Japanese |
| Modalities | Text only | Text |
| Knowledge cutoff | January 2026 | See the model card |
| Access | Free API key | Apache 2.0 weights |
The supported features are chat, tool calling, structured outputs, code generation and streaming. LG also lists EXAONE Tabular, a model for structured data, on the same platform.
How to make your first call
- Get an API key from the EXAONE API Platform at docs.exaone.ai and keep it out of client-side code, as the authentication page advises.
- Export it in your shell with
export EXAONE_API_KEY="YOUR_API_KEY". - Send a request to
https://api.exaone.ai/v1/chat/completionswith the headerAuthorization: Bearer $EXAONE_API_KEYand the modelk-exaone_v2.5. - Read the final answer from
message.contentand the reasoning frommessage.reasoning_content. Reasoning token use appears inusage.reasoning_tokens. - Strip leading whitespace before exact string comparisons, because
contentmay begin with a blank line.
Because the endpoint reference follows the OpenAI format, existing OpenAI SDK code usually needs only a new base URL, key and model name. Set stream: true for server-sent events, and add stream_options.include_usage to receive a final usage chunk.
Reasoning and sampling settings
K-EXAONE 2.5 always thinks before it answers. The API sets enable_thinking to true and does not let you disable it, so you tune depth with reasoning_effort instead. The endpoint page recommends low for fast replies, high for math and coding, and max (the default) when accuracy matters more than speed. The agent guide adds a medium level for general analysis and code writing.
- Sampling. LG recommends
temperature=1.0andtop_p=0.95. Defaults are temperature 1.0, top_p 0.95 and top_k -1 (off). - Multi-turn reasoning. Set
chat_template_kwargs.preserve_thinkingto true for agents, so earlier reasoning stays in the input. When false, only reasoning after the last user message is kept. - Token budget.
max_tokenscounts reasoning, so leave headroom when you use max effort on long inputs.
Using it with Hermes Agent and OpenCode
LG’s Starting with Agents guide gives copy-ready configurations for two command-line agents.
For Hermes Agent, add an exaone provider to ~/.hermes/config.yaml with the API base https://api.exaone.ai/v1, key_env: EXAONE_API_KEY, the chat_completions transport and default_model: k-exaone_v2.5. Then run hermes model to select it, or switch mid-session with /model custom:exaone:k-exaone_v2.5. Our Hermes Agent 0.23 guide covers the latest Hermes release.
For OpenCode, add an openai-compatible exaone provider to opencode.json, declare tool and reasoning support, and set compatibility.reasoningField to reasoning_content so reasoning history passes to the next request. Run it with opencode --standalone. See our OpenCode 2.2 guide for the rest of its features.
LG suggests testing three flows when you wire it into any harness: reasoning streaming, a single tool call that reaches a final answer, and tool execution only after the arguments finish streaming.
Rate limits, errors and limitations
The rate limits page says a 429 error is returned when you exceed your allowance. Each response carries per-second, per-minute and per-day headers, such as x-minute-rate-limit-remaining, plus retry-after when available, so build backoff around those headers rather than fixed numbers. Other errors follow standard codes: 400 for invalid requests, including context overflow, and 401 for a missing or rejected key.
- No image input. If you need vision with a large open model, compare Tencent Hy4 or Kimi K3.1 Pro.
- Korean and English only, compared with ten languages for K-EXAONE 2.0.
- No published end date for free access; the changelog says the service is free currently.
Frequently asked questions
Is the K-EXAONE 2.5 API free?
Yes. LG AI Research says anyone, including developers, startups, companies and research institutions, can use K-EXAONE 2.5 at no cost after getting an API key, and the EXAONE API Platform changelog for November 9, 2026 says the service is currently free of charge. Usage is still subject to rate limits.
What is the K-EXAONE 2.5 model ID and endpoint?
Send POST requests to https://api.exaone.ai/v1/chat/completions with the model set to k-exaone_v2.5 and your key as a Bearer token in the Authorization header. The API is OpenAI-compatible, so OpenAI-style clients can use https://api.exaone.ai/v1 as the base URL.
How large is the K-EXAONE 2.5 context window?
The maximum context length is 524,288 tokens (listed as 512K). Input tokens plus max_tokens, which includes reasoning, must stay within that limit, or the API returns a 400 error.
Can I turn off reasoning in K-EXAONE 2.5?
No. The model always reasons before answering. You control the depth with reasoning_effort; the endpoint reference lists low, high and max, with max as the default, and the agent guide also uses medium.
Are K-EXAONE 2.5 weights available to download?
LG’s announcement covers API access only. As of November 9, 2026, the LGAI-EXAONE Hugging Face organization lists K-EXAONE 2.0 weights under Apache 2.0 but no K-EXAONE 2.5 checkpoint.
Sources
- EXAONE API Platform: K-EXAONE 2.5 endpoint reference
- EXAONE API Platform: Models
- EXAONE API Platform: Starting with Agents
- LG AI Research news: K-EXAONE 2.5 API free release (Korean)
- K-EXAONE 2.0 model card on Hugging Face
Image: LG logo, public domain, via Wikimedia Commons. FreeAnswers is not affiliated with LG.
.svg){kind=link}