01
Overview
Abacus.AI has refreshed its Smaug model line around agentic AI, introducing a three-model family aimed at long-running tool use, coding, automation and multimodal work. The September 10, 2026 announcement names Smaug Agentic, Smaug Flash and Smaug Mini. The important context is that Smaug is not a brand-new name: Abacus released earlier Smaug models in 2024. This 2026 release is a new agent-focused generation built on newer open-weight base models.
02
What is the new Smaug line?
The 2026 Smaug family applies Abacus.AI’s agentic fine-tuning approach to three different base models. Smaug Agentic is built on Moonshot AI’s Kimi K3. Smaug Flash is built on DeepSeek V4 Flash. Smaug Mini is built on Qwen3.8 27B.
Abacus describes the common goal as improving behavior in long-running agent loops: the kinds of workloads where a model repeatedly reasons, calls tools, reads results, edits files or data, and continues over many steps. That positioning matters because strong single-turn benchmark scores do not necessarily translate into stable multi-step agent behavior.
The company says the refreshed models are open-weight and intended to be downloadable and deployable under enterprise control. At the time of this review, Smaug-Agentic has a detailed public Hugging Face model card. Abacus also exposes Smaug Flash in its RouteLLM catalog. Availability should still be checked model by model rather than assuming every artifact has identical hosting, licensing or API options.
03
Smaug Agentic
Smaug-Agentic is an agentic supervised fine-tune of Kimi K3. Its public model card describes a 2.8-trillion-parameter Mixture-of-Experts architecture with 104 billion activated parameters, a 1,048,576-token context window and MoonViT-V2 vision encoder. Abacus says the architecture is unchanged from Kimi K3 and the fine-tune changes behavior rather than the underlying model structure.
That architectural compatibility is useful for deployment. Abacus states that inference stacks that already serve Kimi K3 can serve Smaug-Agentic as a drop-in replacement. The model card points to vLLM, SGLang and TokenSpeed serving paths. It also documents OpenAI-compatible chat-completions behavior and says the model uses preserved thinking history across turns, meaning reasoning content and tool calls should be retained when building multi-turn agent loops.
Smaug-Agentic is not a lightweight local model. The public card describes a 2.8T MoE model and Abacus reports running its evaluations on an 8×B300 deployment. That does not establish a universal minimum hardware requirement, so any consumer-GPU or minimum-VRAM claim would be speculation. Treat it as a self-hostable enterprise-scale model, not as a typical laptop model.
Smaug-Agentic benchmarks
Abacus reports improvements over its cited Kimi K3 baseline on several evaluations, including DeepSWE, LiveBench agentic coding, SciCode, AA-LCR and GPQA Diamond. It also reports a lower extreme tail of reasoning length on two benchmarks and says 113 DeepSWE tasks ran for more than seven hours without infrastructure errors or timeouts.
These are useful launch signals, but they are vendor-run results. The Hugging Face card explicitly distinguishes Smaug-Agentic results that Abacus ran itself from reference scores reproduced from other model releases. They should not be described as independent validation or proof that Smaug-Agentic is universally better than frontier closed models.
04
Smaug Flash
Smaug Flash is the faster, lower-cost member of the new line, based on DeepSeek V4 Flash. Abacus positions it for continuously running agents that work across documents, data systems, APIs and automations rather than only coding.
The company’s open-source page says the fine-tune adapts attention-factor matrices through LoRA-derived deltas while keeping compatibility with the base model’s serving stack and 1M-token context. Abacus reports large gains over its DeepSeek V4 Flash baseline on LiveBench agentic coding, AutomationBench and NL2Repo-Bench. Again, these are first-party benchmark claims.
Unlike the other variants, Smaug Flash also has clear hosted pricing in Abacus RouteLLM at the time of this scan: $0.14 per million input tokens and $0.28 per million output tokens. That price belongs to the Abacus RouteLLM offering; it should not be generalized to every possible Smaug deployment or provider.
05
Smaug Mini
Smaug Mini is based on Qwen3.8 27B and targets smaller multimodal reasoning and tool-use workloads. Abacus says it inherits the base model’s image and video input capabilities and is intended for jobs that inspect media, follow instructions and use a smaller number of tools.
Abacus also says the 27B dense model can fit on a single GPU, but hardware suitability still depends on precision, runtime, context length and workload. A separate minimum-VRAM figure should not be published without model-card or runtime evidence.
The company reports Smaug Mini benchmark leads on IFBench, AutomationBench and JobBench against several compared models. Those are vendor-run claims and should remain labeled accordingly.
06
Smaug Agentic vs Flash vs Mini
The three variants target different operating points. Smaug Agentic is the frontier-scale option for difficult coding and long-horizon agent loops. Smaug Flash is designed around fast, frequently running enterprise and personal-agent workloads. Smaug Mini is the compact multimodal option for smaller reasoning and tool-use jobs.
That does not mean one model is objectively best. The right choice depends on infrastructure, latency, cost, modality and task complexity. Independent cross-model testing is still limited this early in the release cycle.
07
Is Smaug open source?
“Open-weight” is the safer description. Abacus says the refreshed models are downloadable, but licenses inherit constraints from their respective bases. Smaug-Agentic explicitly uses the Kimi K3 License. Open weights do not automatically mean an unrestricted permissive open-source license.
08
Can you self-host Smaug?
Yes in principle, but requirements vary sharply. Smaug-Agentic is explicitly designed to run on the same inference stacks as Kimi K3 and is suitable for enterprise GPU clusters or VPC deployment. Smaug Flash is described as compatible with DeepSeek V4 Flash serving stacks. Smaug Mini is the most plausible single-GPU option according to Abacus, but exact memory requirements should be taken from the final model artifact and runtime configuration rather than guessed.
09
Is there a Smaug API?
Abacus RouteLLM currently lists Smaug Flash with hosted token pricing. A universal public API and one pricing table covering all three Smaug variants was not verified during this scan. Developers should distinguish downloadable model weights from hosted API availability.
10
Why this launch matters
Most open-weight model releases compete on general reasoning or coding scores. The refreshed Smaug line is more specifically aimed at a growing problem: models that look capable in short interactions but become inefficient, indecisive or unstable across long tool-driven workflows. If Abacus’s reported gains reproduce independently, the line could become relevant to developers building coding agents, business automation, personal assistants and enterprise self-hosted AI.
The strongest early advantage is optionality. Teams can choose a large Kimi K3-derived model, a faster DeepSeek-derived model or a compact Qwen-derived multimodal model while keeping the same agent-focused product idea.
11
Important limitations
The September 10 announcement is the launch of the refreshed three-model family, not necessarily the first publication date of every artifact. Smaug-Agentic appeared on Hugging Face before the announcement and was updated close to launch. Benchmark gains are primarily reported by Abacus. Exact public weight availability and license details should be verified separately for Flash and Mini as their final model cards surface. Hosted pricing is currently verified for Smaug Flash through RouteLLM, not the entire family. No universal hardware recommendation is justified.
12
FAQ
What are the three 2026 Smaug models?
Smaug Agentic, Smaug Flash and Smaug Mini.
What is Smaug Agentic based on?
Kimi K3. Abacus says the architecture remains unchanged and the model is behaviorally fine-tuned for long-horizon agentic work.
What is Smaug Flash based on?
DeepSeek V4 Flash.
What is Smaug Mini based on?
Qwen3.8 27B.
Does Smaug Agentic have a 1M-token context window?
Its official Hugging Face card lists a 1,048,576-token context length inherited from Kimi K3.
Can Smaug Agentic run with vLLM?
Abacus documents vLLM, SGLang and TokenSpeed as compatible serving paths because the architecture is unchanged from Kimi K3.
How much does Smaug Flash cost through Abacus?
At scan time, Abacus RouteLLM lists Smaug Flash at $0.14 per million input tokens and $0.28 per million output tokens. Reverify before publication because API pricing can change.
Are the benchmark claims independent?
No. The launch figures discussed here are primarily Abacus-run or vendor-reported results and should be interpreted that way.
Sources
Primary and supporting sources
Facts were rechecked against the linked sources immediately before publication. Pricing, product availability and rollout status can change.