Friendli has made LG AI Research’s K-EXAONE 2.0 750B model available through its Model APIs, adding one of the largest open-weight multilingual models to the company’s serverless catalogue. Friendli recorded the addition in its official changelog on 6 August 2026.
The listing gives developers a managed route to a model with 750 billion total parameters and 37 billion active parameters per token. K-EXAONE 2.0 uses a mixture-of-experts architecture, so only part of the full model participates in each inference step. LG AI Research describes it as an agentic model designed for reasoning, coding, tool use and long-context work.
A very large model without the self-hosting footprint
The practical change is access. LG’s own deployment examples for SGLang and vLLM use two nodes containing eight NVIDIA H200 GPUs each. That is a substantial infrastructure commitment before a team can test the model at useful scale. Friendli’s Model APIs remove the need to provision and operate that stack for ordinary API evaluation and application development.
Friendli’s changelog confirms availability but does not publish a price for this new model in the announcement. Its public pricing page did not yet show K-EXAONE 2.0 750B at the time of review, so buyers should check the Friendli console for the current input, cached-input and output rates before estimating production costs.
The availability is separate from LG’s original model release. LG AI Research published K-EXAONE 2.0 at the end of July; Friendli’s 6 August update is about hosted access through its OpenAI-compatible Model APIs.
Long context and two reasoning modes
K-EXAONE 2.0 supports a context window of 262,144 tokens and has a stated knowledge cut-off of the second quarter of 2025. It covers Korean, English, Spanish, German, Japanese, Vietnamese, French, Italian, Polish and Portuguese.
Applications can use the model in reasoning or non-reasoning mode. LG also documents a preserve-thinking option intended for longer agentic tasks, where retaining reasoning state between turns can help avoid repeatedly reconstructing a plan. Tool calling is supported, which makes the model relevant to coding agents and workflows that need to invoke external functions rather than only return prose.
Under the hood, the model has 78 transformer layers, including two dense layers and 76 sparse layers. Its mixture-of-experts design contains 256 experts and activates eight for a token. These specifications help explain how LG combines a very high total parameter count with a smaller active compute path.
LG released the model weights under the Apache 2.0 licence. That gives organisations a self-hosting route as well as the Friendli option, but the two paths carry different responsibilities. A managed API reduces infrastructure work; it does not answer an organisation’s data-retention, residency or service-level requirements. Those terms need to be checked against Friendli’s current service documentation and contract.
LG’s benchmark claims need real-workload testing
LG reports a score of 68.2 on SWE-bench Verified, 43.8 on Terminal-Bench 2.1 and 94.4 on OpenAI-MRCR. It also reports 89.6 on Ko-LongBench, a Korean long-context benchmark. Those figures are vendor-reported results and do not by themselves establish performance inside a particular coding repository, retrieval system or multilingual support workflow.
The model card also describes multi-token prediction and DSpark speculative decoding, with a claimed three-to-five-times decoding speed-up in suitable deployments. Friendli has not stated in the changelog which of those serving optimisations are exposed through its API, so latency and throughput should be measured directly rather than inferred from the self-hosted configuration.
The evaluation should start with constraints, not scale
The large context window may be useful for repository-scale coding, lengthy document analysis and multi-step agent sessions, but it does not remove normal context-management work. Teams still need to test retrieval quality, instruction adherence, tool-call reliability and the cost of sending very long prompts.
A sensible comparison would keep prompts, tool schemas and decoding settings consistent across K-EXAONE 2.0 and the models already in use. Tests should include malformed tool arguments, long sessions that approach the context limit and representative Korean as well as English tasks. That produces evidence about application behaviour rather than relying on a single aggregate benchmark.
LG cautions that the model can produce false, contradictory, biased or inappropriate responses and that its knowledge is not current beyond the stated cut-off. Applications handling consequential decisions should therefore keep human review, grounding and output validation in place.
For developers already using an OpenAI-compatible client, Friendli’s addition lowers the operational barrier to a serious K-EXAONE 2.0 trial. The unanswered questions are commercial and workload-specific: current pricing, regional performance, sustained throughput and whether the model’s multilingual and agentic strengths hold up on the organisation’s own data.