What is Sage-A1.5?
Sage-A1.5 is an open-weight agent model from the InternScience team at Shanghai AI Laboratory, released on 15 October 2026 with a technical report. It is a 35-billion-parameter mixture-of-experts model built on Alibaba's Qwen3.6-35B-A3B and the lab's earlier Agents-A1, and it reads text and images.
It adds two opt-in modes. Agentic Orchestration lets the model call a spawn_subagent tool to send independent sub-questions to parallel sub-agents, each with its own context and tools. Verification Mode, switched on by a chat-template flag, has the model stop mid-task inside its reasoning to review what has worked and change course.
The weights are under Apache 2.0, with FP8 and Q4_K_M GGUF builds. The native context is 262,144 tokens, extendable to 524,288 with YaRN. The lab offers a free trial through its Intern Discovery token plan; paid prices are not published on the card.
How much does Sage-A1.5 cost?
| Plan | Price | What you get |
|---|---|---|
| Weights | Free | Apache 2.0; BF16, FP8, GGUF |
| Intern Discovery token plan | Free trial | Paid prices not published |
Open weights under Apache 2.0; free trial on Shanghai AI Lab's Intern Discovery token plan; paid API prices not published.
Is Sage-A1.5 worth it?
Sage-A1.5 is more than a fine-tune with a new name: it ships agent behaviour, sub-agent fan-out and in-reasoning self-review, that most open models leave to the harness. On the lab's figures that pays off where it was aimed. It scores 81.7 on BrowseComp (82.8 with orchestration), against 67.9 for its Qwen3.6-35B-A3B base and 75.5 for Agents-A1, and 61.2 on SWE-bench Pro, up 17.6 points on A1 and close to GLM-5.3's 64.6.
The headline wins over far larger models need care. Its "overall best" claims rest heavily on FrontierScience-Research (48.3 against 26.7 for GPT-5.6-sol) and CMT (80.0 against 50.0), benchmarks few outside labs publish, while on established tests it still trails: Kimi-K3 leads BrowseComp at 91.2 and GPT-5.6-sol leads HLE with tools at 64.5 to Sage's 55.1. The team took rivals' numbers from their own reports where available and ran the rest on its own open evaluation code, so the setups are not identical. One tracker has no index score for it yet because there are too few independent results.
The practical design is good. Orchestration is just a tool schema your harness implements, with batch and turn limits (five tasks, 20 turns by default) and sub-agents that return UNKNOWN rather than guess. Verification is a single boolean, though it requires a non-empty system message or the server returns an error. The 256K native context covers long research sessions, and FP8 and GGUF builds bring it within reach of a single GPU.
Against Nanbeige4.3-33B, its nearest open rival for agent work, Sage reads images and has a mixture-of-experts design that is cheaper per token; against its own Qwen base, the gains on search and coding agents are large enough to justify the switch for that work.
Where it earns its place
A data scientist at a 50-person biotech
Run literature and database searches as a self-hosted agent
Register spawn_subagent with search and page-reading tools; the model splits a question into up to five parallel sub-tasks that return evidence and an UNKNOWN status when unsure. Weights stay on-premises under Apache 2.0.
A platform engineer at a mid-size software company
Swap a Qwen3.6-35B-A3B coding agent for something stronger without new hardware
Sage uses the same architecture and tool parsers, so the vLLM command barely changes, and the lab's SWE-bench Pro score rises from 49.5 to 61.2. Re-run your own regression suite before rollout.
An analyst team at a consultancy
Extract figures from scanned reports and cross-check them
It reads images, scoring 91.5 on OmniDocBench in the lab's tests, and Verification Mode makes it review its evidence mid-answer; read reasoning_content to see the self-check.
Before you commit
- Setup
- vllm serve InternScience/Sage-A1.5 with the qwen3 reasoning and qwen3_coder tool parsers, or the SGLang equivalent; add --language-model-only to skip the vision encoder and free cache memory.
- Integrations
- vLLM, SGLang (with MTP speculative decoding), llama.cpp and Ollama via GGUF; OpenAI-compatible tool calling; open evaluation framework in the Agents-A1 GitHub repository.
- Data and privacy
- Apache 2.0, no use restrictions. Data terms for the Intern Discovery hosted trial are not on the model card; self-host for sensitive work.
- Cost at scale
- About 72GB in BF16 for 35.95B parameters, so one 80GB GPU with little cache to spare; the FP8 build halves that, and the Q4_K_M GGUF runs in llama.cpp, Ollama and LM Studio.
- Limits
- 256K native context, 512K only with YaRN; Verification Mode rejects requests without a system message; recommended presence_penalty 1.5; independent results still scarce.
What’s good
- BrowseComp 81.7 and SWE-bench Pro 61.2, well above its Qwen3.6 base on the lab's tests
- Built-in parallel sub-agents and an opt-in self-review mode
- Reads images; 256K native context, 512K with YaRN
- Apache 2.0 with FP8 and GGUF builds
What’s not
- Several of its wins are on benchmarks few others report
- Still well behind Kimi-K3 and GPT-5.6-sol on BrowseComp and HLE
- Hosted pricing not published beyond a free trial
- Orchestration needs your harness to run the sub-agents
Best for
Teams building self-hosted research or browsing agents who want sub-agent fan-out and self-checking in the model rather than in custom harness code.
Sage-A1.5 alternatives
Sage-A1.5: common questions
- Is Sage-A1.5 worth it?
- Sage-A1.5 turns a Qwen base into a credible open research agent with sub-agents built in, though its big wins are on benchmarks few others report. We rate it 3.9 out of 5.
- How much does Sage-A1.5 cost?
- Open weights under Apache 2.0; free trial on Shanghai AI Lab's Intern Discovery token plan; paid API prices not published.
- What are the downsides of Sage-A1.5?
- Several of its wins are on benchmarks few others report; still well behind Kimi-K3 and GPT-5.6-sol on BrowseComp and HLE; hosted pricing not published beyond a free trial; orchestration needs your harness to run the sub-agents.
- Who is Sage-A1.5 best for?
- Teams building self-hosted research or browsing agents who want sub-agent fan-out and self-checking in the model rather than in custom harness code.
- How much does Sage-A1.5 cost for a team or in production?
- About 72GB in BF16 for 35.95B parameters, so one 80GB GPU with little cache to spare; the FP8 build halves that, and the Q4_K_M GGUF runs in llama.cpp, Ollama and LM Studio.
- Is Sage-A1.5 safe for company data?
- Apache 2.0, no use restrictions. Data terms for the Intern Discovery hosted trial are not on the model card; self-host for sensitive work.
- What are the best Sage-A1.5 alternatives?
- The closest alternatives we have reviewed are Nanbeige4.3-33B, Qwen3.8-27B, Kimi K3 and GLM-5.3.
Answers checked Oct 31, 2026.