Abacus.AI has made Smaug Flash available through its RouteLLM API at $0.10 per million input tokens and $0.40 per million output tokens. CEO Bindu Reddy announced the open-weights fine-tune in a post on X.
The company describes Smaug Flash as a fine-tune of DeepSeek V4 Flash for agentic and coding tasks. Its RouteLLM listing gives it a 1 million-token context window and a LiveBench score of 77.4.
For personal-agent use, Abacus.AI says Smaug Flash is fine-tuned on DeepSeek Flash and can connect to WhatsApp, Telegram and Slack. The company says the model can be downloaded through Hugging Face.
Abacus.AI announced the broader Smaug line on September 10, 2026. The line includes Smaug Agentic, Smaug Flash and Smaug Mini, according to the company’s release.
On its open-source page, Abacus.AI says Smaug Flash improved on base DeepSeek V4 Flash by 14.3 points on LiveBench agentic coding, 13.7 points on AutomationBench and 19 points on NL2Repo-Bench. The company also says the model led Claude Sonnet 5 on each benchmark it compared.
Reddy is Abacus.AI’s CEO. She co-founded the San Francisco company with CTO Arvind Sundararajan; both are veterans of Google, Amazon and Uber. Abacus.AI began in 2019 as RealityEngines.AI, rebranded in July 2020 and raised a $13 million Series A round. It later raised a $50 million Series C round in late 2021, bringing total funding to roughly $90 million at the time.
The company describes its platform as a SuperAssistant and agentic platform for white-collar work. It says its multi-model router provides access to more than 100 LLMs, alongside its Smaug models.
