Abacus.AI Smaug

Smaug is Abacus.AI's line of three open-weight language models: Smaug Agentic, Smaug Flash and Smaug Mini. They are fine-tuned from Kimi K3, DeepSeek V4 Flash and Qwen3.8 27B for long-running agentic and coding loops. The line is for enterprise AI teams that want stronger agent performance they can self-host in their own VPC, or call cheaply through Abacus.AI's RouteLLM API.

Abacus.AI launched the Smaug line on 10 September 2026: three open-weight models fine-tuned for enterprise agentic workloads and published on Hugging Face. Smaug is a fine-tuning technique that Abacus.AI says works on any open-source base model. It combines human-curated agentic traces with synthetic data and masks reasoning tokens from the training loss to preserve the base model's reasoning. Smaug Agentic is built on Moonshot AI's Kimi K3: a 2.8-trillion-parameter mixture-of-experts model with 104 billion active parameters, a 1,048,576-token context window and a vision encoder, released under the inherited Kimi K3 License. It runs on vLLM, SGLang and TokenSpeed as a drop-in for Kimi K3 serving stacks and was evaluated on 8 B300 GPUs. It scores 69.9 on DeepSWE against 67.5 for the base, 64.6 against 62.2 on LiveBench agentic coding, and 94.1 against 93.5 on GPQA Diamond. Abacus.AI says it sustained a median of 78 agent steps across 113 DeepSWE tasks over seven hours with zero infrastructure errors. Smaug Flash is a 304B fine-tune of DeepSeek-V4-Flash-0731 with 1M context, released under MIT. It lifts LiveBench agentic coding from 46.8 to 61.1 and AutomationBench from 25.1 to 38.83, is aimed at personal agents on WhatsApp, Telegram and Slack, and costs $0.10 per million input tokens and $0.40 per million output tokens on RouteLLM. Smaug Mini is a 27B Apache 2.0 fine-tune of Qwen3.8-27B with 262K context, for multimodal chatbots and smaller reasoning jobs; it lifts JobBench from 33.4 to 50.5. Abacus.AI claims 15-20% better performance on long agent loops at no extra cost, and prices 10-100 times below Anthropic and OpenAI frontier models. Abacus.AI is based in San Francisco, founded by Bindu Reddy and Arvind Sundararajan, and backed by Coatue, Tiger Global, Index Ventures and Khosla Ventures. It reports more than 3 million users. It previously used the Smaug name for Smaug-72B in 2024.

Category

AI Models & APIs

Pricing

Free, Subscription, Usage-based

Target Market

CTOs, Heads of AI, ML Platform Engineers, Enterprise Developers

Deployment

Open-source, Self-hosted, API-based

Headquarters

San Francisco, United States

Customers

Abacus.AI reports 3M+ users across professionals, small businesses and enterprises, including dozens of the Fortune 500 (company-wide, not Smaug-specific)

• ✓
  Smaug Agentic (Kimi K3 base)

  A 2.8T-parameter MoE model with 104B active parameters and 1M context, tuned for complex long-running coding loops. Abacus.AI positions it as a self-hostable replacement for Opus-class closed models.

• ✓
  Smaug Flash (DeepSeek V4 Flash base)

  A 304B MIT-licensed model with large gains on agentic coding and AutomationBench over its base. It is priced at $0.10/$0.40 per million tokens on RouteLLM.

• ✓
  Smaug Mini (Qwen3.8 27B base)

  A 27B Apache 2.0 multimodal model with 262K context for chatbots and smaller reasoning tasks. It gives agentic gains in a size that is practical to self-host.

• ✓
  Agent-trace fine-tuning method

  Combines human-curated agentic traces with synthetic data and masks reasoning tokens from the loss. The aim is to improve tool-using behaviour without degrading the base model's reasoning.

• ✓
  Drop-in serving compatibility

  The models run on vLLM, SGLang and TokenSpeed using the same serving recipes as their base models, so existing inference infrastructure can switch with little change.

• ✓
  OpenAI-compatible RouteLLM API

  Smaug Flash is served through RouteLLM's OpenAI-compatible endpoint alongside 160+ other models. Teams can compare it with closed models without new integration work.

• •
  Replacing a closed model in a coding agent

  An engineering platform team swaps an Opus-class closed model for self-hosted Smaug Agentic to cut per-token cost on multi-hour autonomous coding runs.

• •
  Messaging-channel personal agents

  A company builds WhatsApp and Slack assistants on Smaug Flash through RouteLLM, paying $0.10 per million input tokens for high-volume conversational agent traffic.

• •
  In-VPC regulated agent workloads

  A financial-services firm hosts Smaug models inside its own cloud VPC, so proprietary data used by automation agents never reaches a third-party model provider.

• •
  Compact multimodal support bot

  A support organisation deploys the 27B Apache 2.0 Smaug Mini with 262K context to handle image-plus-text customer queries on modest self-hosted GPU capacity.

All three models can be downloaded free, so self-hosting costs are infrastructure costs: Smaug Agentic was evaluated on 8 B300 GPUs. The managed route requires a ChatLLM Teams subscription ($7 for the first month, then $10 a month) for RouteLLM credentials, plus per-token billing; Smaug Flash is $0.10 per million input and $0.40 per million output tokens. Smaug Mini and Smaug Agentic did not appear in the RouteLLM catalogue at research time. An independent pricing guide criticises the ChatLLM credit system as opaque.

This page was written from 12 sources, 10 on domains other than abacus.ai.

1. 1.abacus.ai — smaugvendor
2. 2.abacus.ai — aboutvendor
3. 3.prnewswire.com — abacusai launches the smaug line of open weight models optim
4. 4.unite.ai — abacus ai releases three open weight smaug models for agenti
5. 5.techedgeai.com — abacus ai smaug targets cheaper enterprise ai agents
6. 6.huggingface.co — Smaug Agentic
7. 7.huggingface.co — Smaug Flash
8. 8.huggingface.co — Smaug Mini
9. 9.huggingface.co — abacusai
10. 10.routellm-apis.abacus.ai — routellm-apis.abacus.ai
11. 11.eesel.ai — abacus ai pricing
12. 12.hn.algolia.com — search