01
Overview
Smaug-Agentic is Abacus.AI’s Kimi K3-derived model for long-horizon coding and tool-use loops. Its strongest deployment advantage is compatibility: Abacus says the fine-tune does not change Kimi K3’s architecture, tokenizer, context window or vision encoder, so inference stacks capable of serving the base model can serve Smaug-Agentic as a drop-in replacement.
That does not make it a casual laptop model. The official model card describes a 2.8T-parameter Mixture-of-Experts model with 104B activated parameters and a 1,048,576-token context window. This guide therefore treats “run locally” as self-managed infrastructure, not as a promise that a typical desktop GPU can load the model.
02
Official model details
The Hugging Face card lists:
- Base model: moonshotai/Kimi-K3
- Architecture: Mixture-of-Experts
- Total parameters: 2.8T
- Activated parameters: 104B
- Context length: 1,048,576 tokens
- Modality: text and image
- Vision encoder: MoonViT-V2
- Quantization in the model summary: MXFP4 weights / MXFP8 activations
- License: inherited Kimi K3 License
Abacus says every architectural parameter remains unchanged from the base model. The tuning is behavioral, using multi-turn tool-using coding trajectories.
03
Before you deploy
The first question is not “which consumer GPU?” It is whether your existing infrastructure can serve Kimi K3. If it can, Abacus says Smaug-Agentic can use the same inference path. If it cannot, the model card does not establish a small minimum-VRAM shortcut.
Abacus’s own published evaluation used a dedicated 8×B300 deployment. That is an evaluation configuration, not an official minimum hardware requirement. Do not convert it into a claim that eight B300s are required, and do not publish a smaller guessed requirement.
04
vLLM path
The official Hugging Face integration page provides a vLLM serving path and shows the model being served directly from the Hugging Face repository. The high-level workflow is:
- Install a compatible vLLM release.
- Start a vLLM server with abacusai/Smaug-Agentic as the model repository.
- Send OpenAI-compatible chat-completions requests to the local server.
- Preserve model reasoning and tool-call fields across turns when building an agent loop.
The exact vLLM version, distributed-GPU flags, tensor parallelism and memory settings should follow current Kimi K3 serving recipes and the hardware actually available. This draft intentionally does not invent one configuration that will work on every cluster.
05
SGLang path
Abacus also lists SGLang as a supported serving stack because the architecture matches Kimi K3. The same principle applies: use the current Kimi K3 cookbook for distributed serving and substitute the Smaug-Agentic repository where appropriate.
SGLang can be attractive for agent-serving environments that already use an OpenAI-compatible frontend and distributed inference. But compatibility at the model level does not guarantee identical throughput or memory behavior on every cluster, so benchmark your actual deployment.
06
TokenSpeed path
The model card additionally points to TokenSpeed recipes. This is another official compatibility path, but it is less likely to be the first choice for readers already standardized on vLLM or SGLang. Keep it in the deployment-options section rather than creating a separate thin article.
07
Preserving reasoning history matters
Smaug-Agentic always uses thinking according to its model card. Reasoning effort is configured with low, high or max, with max documented as the default in the release card.
For multi-turn conversations and tool calls, Abacus specifically says the complete assistant message returned by the API should be passed back into the next messages array. That includes reasoning_content and tool_calls, not only visible content.
This matters because the model was trained in a preserved-thinking-history mode. Dropping the reasoning/tool state between turns can change the behavior of a long-running agent and undermine the type of workflow the fine-tune targets.
08
OpenAI-compatible agent frameworks
The model card says Smaug-Agentic works with agent frameworks that speak the OpenAI chat-completions contract. Abacus also notes a practical compatibility issue: some OpenAI-compatible servers reject unrelated provider-specific fields when full historical messages are echoed back.
Its guidance is to keep reasoning_content and tool_calls but strip unsupported fields such as provider-specific metadata when a server rejects them. Treat this as implementation guidance from Abacus rather than a universal OpenAI API rule.
09
Sampling settings from the model card
Abacus documents temperature 1.0 for its published evaluations, with top-p 0.95 for single-step tasks and top-p 1.0 for agentic tasks. These settings describe the release evaluation and recommended sampling behavior; they are not proof that other settings are invalid.
Reasoning effort can materially affect latency and token use, so production testing should compare quality and cost at more than one effort level rather than defaulting every request to maximum effort without measurement.
10
Can Smaug-Agentic run on one GPU?
The official evidence used here does not establish that. The model is extremely large even though only part of the MoE activates per token. “104B activated parameters” does not mean only 104B parameters need to be stored. Storage, memory placement, KV cache, context length, runtime and precision all matter.
If community quantizations or alternative serving recipes appear, label them as community artifacts and test them independently. Do not present unofficial GGUF conversions or estimated VRAM tables as Abacus requirements.
11
Is Smaug-Agentic actually open source?
Use “open-weight” for precision. The weights are available through Hugging Face, but the model inherits the Kimi K3 License. Anyone deploying commercially should read those terms rather than assuming an MIT or Apache-style license.
12
What should you benchmark after deployment?
For agentic workloads, measure more than tokens per second. Useful production metrics include task success rate, number of agent steps, repeated tool calls, failed tool calls, total reasoning tokens, wall-clock completion time, context growth and infrastructure stability.
Abacus’s central claim is improved long-horizon behavior. The best validation is therefore a paired test against Kimi K3 using the same agent scaffold, prompts, tools, infrastructure and stopping rules.
13
FAQ
Does Smaug-Agentic support vLLM?
Yes. Abacus’s official Hugging Face model card documents vLLM as a serving path and says the unchanged Kimi K3 architecture makes existing K3 inference stacks compatible.
Does it support SGLang?
Yes. SGLang is another documented serving path.
What is the context window?
The model card lists 1,048,576 tokens.
What GPU do I need?
No universal minimum GPU or VRAM requirement is established by the official model card. Abacus’s benchmark deployment used 8×B300, but that is not stated as the minimum deployment configuration.
Can I use an OpenAI-compatible agent framework?
Abacus says yes. Preserve reasoning_content and tool_calls across turns for its documented multi-turn behavior.
What license does it use?
The Kimi K3 License inherited from the base model.
Sources
Primary and supporting sources
Facts were rechecked against the linked sources immediately before publication. Pricing, product availability and rollout status can change.