InternScience/Sage-A1.5-FP8

Sage-A1.5

🏠 Homepage | Technical Report | 🎁 Free Trial

Hugging Face | Github | ModelScope

This repository contains model weights and configuration files for Sage-A1.5 in the Hugging Face Transformers format.

These artifacts are compatible with Hugging Face Transformers, vLLM, SGLang, etc.

---

Sage‑A1.5 is a 35B Mixture‑of‑Experts agentic model from Shanghai AI Laboratory, built to advance long-horizon scientific problem solving through adaptive capability composition. Building upon Agents‑A1, Sage‑A1.5 learns not only to solve individual tasks, but also to organize work and revise its approach during execution, highlighting three key capabilities: Agentic Orchestration, Verification, and Open Scientific Discovery.

First, agentic orchestration extends single-agent execution through task delegation and parallel collaboration among subagents with isolated contexts, enabling complex tasks to be decomposed and addressed concurrently while reducing interference across concurrent subagents. Second, agentic verification strengthens long-horizon reliability through progress checking, evidence validation, and reflective strategy revision, allowing Sage‑A1.5 to recognize unproductive search paths and pursue alternative approaches. Third, open scientific discovery extends agentic capabilities beyond answering predefined questions to conducting exploratory research, connecting scientific inquiry with interactive analysis and inspectable research artifacts. We hope Sage‑A1.5 provides the community with an accessible foundation for open scientific discovery, offering a practical path from long-horizon task execution to collaborative, verifiable, and inspectable scientific exploration.

Highlights

• Agentic Reasoning: Sage-A1.5 excels at decomposing complex tasks into executable sub-steps, planning ahead, and adapting its strategy based on intermediate results.
• Tool Use: Natively supports function calling and tool integration, enabling seamless interaction with APIs, code interpreters, search engines, and other external tools.
• Scientific and Professional Reasoning: Handles tool-integrated scientific reasoning and professional knowledge question answering.
• Instruction Following: Precisely follows detailed, multi-constraint instructions across diverse domains.
• Agentic Orchestration: Orchestrates parallel investigations across subagents with independent contexts, enabling them to challenge assumptions, refine conclusions through evidence, and synthesize their findings into comprehensive results.
• Verification Mode: An opt-in self-review capability that lives inside the model's own chain of thought — the model pauses mid-trajectory to review what it has established and what has failed, then acts on a revised direction in the same turn, with no external critic.

We welcome developers and enterprises to integrate and try Sage-A1.5 and share their feedback.

Performance

We evaluate Sage-A1.5 in real-world agentic and research-oriented workflows across six directions — long-horizon search, engineering tasks, instruction following, scientific research, general agentic tasks, and scientific agentic tasks. Despite operating in the ~35B model class, Sage-A1.5 delivers highly competitive performance against frontier-scale systems such as Kimi-K3, GLM-5.3, and GPT-5.6-sol. It achieves overall SOTA on XBench-DS-2510 (92.0), IFBench (81.6), PinchBench (90.0), FrontierScience-Olympiad (81.0), FrontierScience-Research (48.3), CMT (80.0), and OmniDocBench (91.5), and reaches 95.1 on GAIA — higher than every larger-scale model we compare against. Among comparable ~35B models it ranks first on a broad range of tasks, including BrowseComp (81.7), SWE-bench Pro (61.2), MLE-Lite (65.2), SciCode (44.3, tied with Agents-A1), AA-LCR (81.0), HLE w/ tools (55.1), VitaBench (45.4), and ChemCOTBench (66.5). Compared with A1, it improves substantially on engineering and agentic workloads (SWE-bench Pro +17.6, SWE-bench Multilingual +14.0, PinchBench +13.1, MLE-Lite +21.3). These results show that Sage-A1.5 combines strong long-horizon search ability, robust scientific reasoning, and reliable instruction following, establishing it as a highly capable and efficient agentic model that narrows the gap with much larger frontier models.

🥇 Overall SOTA    🟢 Best Among Comparable Models (~35B)

Benchmark	📏 Comparable Models (~35B)			🚀 Larger-scale Models
	Sage-A1.5	Agents-A1	Ornith-1.5	Kimi-K3	GLM-5.3	GPT-5.6-sol
🔍 Long-horizon Search
GAIA	95.1	96.0 🥇	77.6	92.2	92.2	88.3
BrowseComp	81.7 🟢 w/o Agentic Orchestration 82.8 w/ Agentic Orchestration	75.5	67.6	91.2 🥇	76.2	90.4
DeepResearch-Bench2	43.7 w/o Agentic Orchestration 47.5 w/ Agentic Orchestration	37.5	48.6 🟢	52.4	55.7	57.9 🥇
XBench-DS-2510	92.0 🥇	86.0	80.0	85.0	84.0	90.0
⚙️ Engineering Tasks
MLE-Lite	65.2 🟢	43.9	39.4	72.7	72.7	77.3 🥇
SciCode	44.3 🟢	44.3	39.5	58.7	59.0 🥇	56.1
SWE-bench Pro	61.2 🟢	43.6	59.6	63.3	64.6	64.6 🥇
SWE-bench Multilingual	72.7 🟢	58.7	71.4	80.8	81.3 🥇	74.1
📋 Instruction Following
IFBench	81.6 🥇	80.6	44.6	72.1	53.4	72.7
AA-LCR	81.0 🟢	74.2	73.2	88.7 🥇	79.7	84.0
🧪 Scientific Research
FrontierScience-Olympiad	81.0 🥇	79.0	70.0	69.0	67.0	73.0
FrontierScience-Research	48.3 🥇	40.0	6.7	21.7	16.7	26.7
HiPhO	45.3	46.4 🥇	39.5	42.4	39.7	46.2
HLE w/ tools	55.1 🟢	47.6	33.4	56.0	62.5	64.5 🥇
CMT	80.0 🥇	72.0	26.0	34.0	40.0	50.0
CritPt	20.0 🟢	4.3	2.9	23.4	19.1	32.3 🥇
🤖 General Agentic Tasks
Agents' Last Exam (ALE-CLI)	19.1 🟢	10.2	14.3	27.6	28.5	28.6 🥇
VitaBench	45.4 🟢	38.7	44.6	59.0 🥇	58.7	46.8
PinchBench	90.0 🥇	76.9	88.6	88.2	86.6	87.3
🔬 Scientific Agentic Tasks
ChemCOTBench	66.5 🟢	64.0	53.8	82.5 🥇	78.3	77.9
MatTools	42.6	47.1 🟢	38.6	24.3	53.2 🥇	45.6
OmniDocBench	91.5 🥇	89.5	89.6	91.1	-	85.8

SWE-Bench Series: We use the mini-SWE-agent harness with a temperature of 1.0, top-p of 0.95, a 200K-token context window, and automatic context compaction enabled. All evaluations are conducted in a fully offline environment without internet access.

GLM-5.3 on OmniDocBench: GLM-5.3 is a text-only (unimodal) model without multimodal input support, so its results on OmniDocBench, a multimodal benchmark, are not reported.

Usage

SGLang

SGLang is a fast serving framework for large language models and vision language models.

Install SGLang with uv:

uv venv --python 3.12 --seed --managed-python
source .venv/bin/activate

uv pip install sglang

See its documentation for more details.

The following commands create API endpoints at http://localhost:8000/v1:

• Standard Version:

  python -m sglang.launch_server \
    --model-path InternScience/Sage-A1.5 \
    --port 8000 \
    --mem-fraction-static 0.8 \
    --context-length 262144 \
    --reasoning-parser qwen3 \
    --tool-call-parser qwen3_coder
  

• Multi-Token Prediction (MTP):

  python -m sglang.launch_server \
    --model-path InternScience/Sage-A1.5 \
    --port 8000 \
    --mem-fraction-static 0.8 \
    --context-length 262144 \
    --reasoning-parser qwen3 \
    --tool-call-parser qwen3_coder \
    --speculative-algorithm EAGLE \
    --speculative-num-steps 3 \
    --speculative-eagle-topk 1 \
    --speculative-num-draft-tokens 4
  

For ultra long context:

Sage-A1.5 natively supports a 256K context window. To extend it to 512K, serve the model with YaRN (factor: 2.0) enabled by overriding the RoPE parameters:

SGLANG_ALLOW_OVERWRITE_LONGER_CONTEXT_LEN=1 \
python -m sglang.launch_server \
  --model-path InternScience/Sage-A1.5 \
  --port 8000 \
  --context-length 524288 \
  --json-model-override-args '{
      "text_config": {
          "rope_parameters": {
              "mrope_interleaved": true,
              "mrope_section": [11, 11, 10],
              "rope_type": "yarn",
              "rope_theta": 10000000,
              "partial_rotary_factor": 0.25,
              "factor": 2.0,
              "original_max_position_embeddings": 262144
          }
      }
  }' \
  --reasoning-parser qwen3 \
  --tool-call-parser qwen3_coder \
  --mem-fraction-static 0.88

vLLM

vLLM is a high-throughput and memory-efficient inference and serving engine for LLMs.

Install vLLM from the main branch via uv:

uv venv --python 3.12 --seed --managed-python
source .venv/bin/activate

uv pip install vllm --torch-backend=auto

See its documentation for more details.

The following commands create API endpoints at http://localhost:8000/v1:

• Standard Version (1 GPUs, 262K context):

  vllm serve InternScience/Sage-A1.5 \
    --port 8000 \
    --max-model-len 262144 \
    --reasoning-parser qwen3 \
    --enable-auto-tool-choice \
    --tool-call-parser qwen3_coder
  

• Text-Only (skips vision encoder to free KV cache memory):

  vllm serve InternScience/Sage-A1.5 \
    --port 8000 \
    --tensor-parallel-size 1 \
    --max-model-len 262144 \
    --reasoning-parser qwen3 \
    --language-model-only
  

• Multi-Token Prediction (MTP):

  vllm serve InternScience/Sage-A1.5 \
    --port 8000 \
    --max-model-len 262144 \
    --reasoning-parser qwen3 \
    --enable-auto-tool-choice \
    --tool-call-parser qwen3_coder \
    --speculative-config '{"method":"mtp","num_speculative_tokens":3,"moe_backend":"triton"}'
  

For ultra long context:

Sage-A1.5 natively supports a 256K context window. To extend it to 512K, serve the model with YaRN (factor: 2.0) enabled via --hf-overrides (vLLM no longer supports the old --rope-scaling flag):

VLLM_ALLOW_LONG_MAX_MODEL_LEN=1 \
vllm serve InternScience/Sage-A1.5 \
  --port 8000 \
  --max-model-len 524288 \
  --hf-overrides '{
    "text_config": {
      "rope_parameters": {
        "rope_type": "yarn",
        "rope_theta": 10000000,
        "partial_rotary_factor": 0.25,
        "factor": 2.0,
        "original_max_position_embeddings": 262144,
        "mrope_interleaved": true,
        "mrope_section": [11, 11, 10]
      }
    }
  }' \
  --reasoning-parser qwen3 \
  --enable-auto-tool-choice \
  --tool-call-parser qwen3_coder

Advanced Modes

Sage-A1.5 ships two opt-in modes that build on the standard serving setup above. Both are off by default: Agentic Orchestration Mode is turned on by registering a tool, Verification Mode by a chat-template flag.

Agentic Orchestration

Agentic Orchestration Mode (see Highlights) lets the model fan out independent sub-questions to sub-agents that run in parallel, each with its own context window and its own tool access. It is enabled by exposing a spawn_subagent tool to the model: the model calls it whenever a task decomposes into independent investigations, and your harness runs each sub-task and returns the structured results.

Sub-agents are not restricted to any fixed tool set — they can be given whatever tools your deployment registers (search, page reading, code execution, retrieval, internal APIs, …). By default they inherit the main agent's tools; set subagent_tools if you want them to have a different, for instance narrower, set.

Register the following tool schema alongside your other tools:

SPAWN_SUBAGENT_SCHEMA = {
    "type": "function",
    "function": {
        "name": "spawn_subagent",
        "description": (
            "Dispatch one or more sub-agents IN PARALLEL to investigate independent "
            "sub-questions. Each sub-agent has access to google_search and read_page "
            "but CANNOT see your conversation. You MUST give each sub-agent everything "
            "it needs via the `task`, `context`, and `expected_output` fields. Sub-agents "
            "return structured JSON: {answer, status, evidence, unresolved_issues}."
        ),
        "parameters": {
            "type": "object",
            "properties": {
                "tasks": {
                    "type": "array",
                    "minItems": 1,
                    "description": "List of independent sub-tasks to dispatch in parallel.",
                    "items": {
                        "type": "object",
                        "properties": {
                            "task": {
                                "type": "string",
                                "description": "Self-contained sub-question. The sub-agent has no other context.",
                            },
                            "context": {
                                "type": "string",
                                "description": "Background facts and constraints the sub-agent needs.",
                            },
                            "expected_output": {
                                "type": "string",
                                "description": "Form of the desired answer (e.g. 'a year YYYY', 'a person name').",
                            },
                        },
                        "required": ["task", "context", "expected_output"],
                    },
                },
            },
            "required": ["tasks"],
        },
    },
}

Because a sub-agent cannot see the main conversation, every task must be self-contained — task, context, and expected_output are all required. Each sub-agent returns one compact JSON object {answer, status, evidence, unresolved_issues}, and the batch is handed back to the main model in input order.

Notes:

• Keep each sub-task independent. Tasks in one batch run concurrently and cannot see each other's results, so anything sequential belongs in separate spawn_subagent calls.
• expected_output acts as a format constraint on the sub-agent's answer (e.g. 'a year YYYY', 'a person name'), which keeps results short and easy to merge.
• A sub-agent that cannot resolve its task returns "UNKNOWN" with a status of unknown or insufficient_evidence rather than guessing — check status before using an answer.
• Batch size and per-sub-agent turn budget are harness-side settings (max_tasks_per_spawn, default 5; subagent_max_turns, default 20). Oversized batches are rejected with an error message instead of being truncated.
• The schema's description above names google_search and read_page because that is the reference setup. If you give sub-agents a different tool set, update that sentence to match — it is what tells the main model what its sub-agents can actually do.

Verification Mode

Verification Mode (see Highlights) is opt-in and off by default. Enable it per request by passing verification through chat_template_kwargs; the chat template then appends the verification protocol trigger to the end of your system message.

• OpenAI Python client (via extra_body):

  from openai import OpenAI
  
  client = OpenAI(base_url="http://localhost:8000/v1", api_key="EMPTY")
  
  response = client.chat.completions.create(
      model="InternScience/Sage-A1.5-FP8",
      messages=[
          {"role": "system", "content": "You are a helpful assistant."},
          {"role": "user", "content": "Find the population of Beijing and cross-check it."},
      ],
      tools=TOOLS,
      extra_body={"chat_template_kwargs": {"verification": True}},
  )
  
  message = response.choices[0].message
  print(message.reasoning_content)  # self-review happens inside the chain of thought
  print(message.content)
  

• curl:

  curl http://localhost:8000/v1/chat/completions \
    -H 'Content-Type: application/json' \
    -d '{
      "model": "InternScience/Sage-A1.5-FP8",
      "messages": [
        {"role": "system", "content": "You are a helpful assistant."},
        {"role": "user", "content": "Find the population of Shanghai and cross-check it."}
      ],
      "chat_template_kwargs": {"verification": true}
    }'
  

Notes:

• A non-empty system message is required. Because the trigger is appended to the end of the system content, a request with verification: true and no system message (or a whitespace-only one) is rejected with HTTP 400.
• Pass a JSON boolean, not a string: {"verification": true}, not {"verification": "true"}.
• Omitting chat_template_kwargs is equivalent to {"verification": false}, which renders the prompt unchanged.
• Verification Mode is independent of tool use — it works with or without tools.
• The self-review is emitted as reasoning content, so read reasoning_content in addition to content.

Recommended Sampling Parameters

For the best generation quality, we recommend the following sampling parameters:

• temperature: 1.0
• top_p: 0.95
• top_k: 20
• min_p: 0.0
• presence_penalty: 1.5
• repetition_penalty: 1.0

Agent Capability Evaluation

To provide the community with a unified agent evaluation codebase for fair comparison, we have also open-sourced an evaluation framework for assessing agentic models across core capabilities, including tool use and multi-step reasoning. The evaluation code is included in the Sage-A1/evaluation of this repository.

We use this framework to evaluate the released model under a standardized and reproducible setting. Specifically, the model is tested on a set of agent-oriented tasks that require it to understand user goals, decompose complex instructions, interact with tools or environments when necessary, and produce final results. The evaluation results reported in Model Card are generated using the open-source framework above, so that users can reproduce the experiments, compare other models under the same protocol, and further extend the benchmark for new agent scenarios. (Note that: To ensure a fair comparison, we report the benchmark results from their original technical reports. If a model does not report the corresponding benchmark results, we evaluate it using the same evaluation protocol as our model.)

For detailed evaluation scripts, task definitions, metrics, and reproduction instructions, please refer to the evaluation codebase.

Citation

If you find our work helpful, feel free to give us a cite.

@misc{lou2026sagea15longhorizonexecutionscientific,
      title={Sage-A1.5: From Long-Horizon Execution to Scientific Exploration with Capability Composition}, 
      author={Sage-A1.5 Team},
      year={2026},
      eprint={2610.28578},
      archivePrefix={arXiv},
      primaryClass={cs.AI},
      url={https://arxiv.org/abs/2610.28578}, 
}