Sage-A1.5
InternScience's ~35B MoE agent model (3B active) scoring 81.7 on BrowseComp and 61.2 on SWE-Bench Pro, with FP8 and GGUF builds.
This entry is marked for review. Check the original sources before relying on its details.
Overview
Sage-A1.5 is the successor to Agents-A1 from the InternScience Hugging Face organization, described in a paper (arXiv 2610.28578) and reported on October 21, 2026. It is a mixture-of-experts agent model with about 35B parameters and 3B active per token (the Hugging Face collection lists 36B), plus an orchestration layer for subagents and a verification protocol. Reported scores are BrowseComp 81.7 (85.5 with verification), SWE-Bench Pro 61.2 versus 43.6 for Agents-A1, and GAIA 95.1; Kimi-K3 and GPT-5.6-sol still lead on BrowseComp. FP8 and GGUF (Q4_K_M to F16) builds are published. License and base model are not stated on the pages reviewed.
This is a sourced directory entry. Features and capabilities are reported in the linked sources; this entry does not document independent Aymorphic testing.
Key features
- ~35B MoE with 3B active parameters per token
- BrowseComp 81.7 (85.5 with verification); SWE-Bench Pro 61.2; GAIA 95.1
- Subagent orchestration layer and verification protocol
- FP8 and GGUF builds (Q4_K_M, Q6_K, Q8_0, F16)
- Deployment targets listed: NVIDIA A100-B200, AMD MI300X, Huawei Ascend, Apple M-series
Use cases
- Local deep-research agents on a single accelerator
- Agentic coding on SWE-Bench Pro-style tasks
- Subagent-orchestrated scientific exploration workflows
Pricing
Open weights on Hugging Face; license unknown.
Pricing reflects the source record and may change. Confirm current terms with the provider.
Dates and sources
- Released
- 21 Oct 2026
- Record updated
- 28 Oct 2026
- Record last checked
- 28 Oct 2026