--- library_name: gguf base_model: InternScience/Sage-A1.5 license: apache-2.0 pipeline_tag: text-generation tags: - gguf - moe - vlm - vision - agentic - code - tool-use quantized_by: InternScience base_model_relation: quantized --- # Sage-A1.5 GGUF (Q8_0)
🏠 Homepage | Technical Report Technical Report | 🎁 Free Trial
Hugging Face Hugging Face | GitHub Github | Model Scope ModelScope
> [!Note] > This repository contains **GGUF** model weights for Sage-A1.5, quantized to **Q8_0**, for deployment with [llama.cpp](https://github.com/ggml-org/llama.cpp), Ollama, LM Studio, and other GGUF-compatible runtimes. > > For the original Hugging Face Transformers weights (compatible with Transformers, vLLM, SGLang, etc.), see [InternScience/Sage-A1.5](https://huggingface.co/InternScience/Sage-A1.5). --- **Sage‑A1.5** is a 35B Mixture‑of‑Experts agentic model based on **[Qwen3.6](https://huggingface.co/Qwen/Qwen3.6-35B-A3B)** and **[Agents-A1](https://huggingface.co/InternScience/Agents-A1)** 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. ![Sage-A1 Benchmark Overview](./figures/Sage-A1p5-benchmark.png) ## 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 β€” 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 β€” 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 β€” 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.

πŸ₯‡ Overall SOTA    🟒 Best Among Comparable Models (~35B)

Benchmark πŸ“ Comparable Models (~35B) πŸš€ Larger-scale Models
Sage-A1.5 Agents-A1 Qwen3.6-35B-A3B Ornith-1.5 Kimi-K3 GLM-5.3 GPT-5.6-sol
πŸ” Long-horizon Search
GAIA 95.1 96.0 πŸ₯‡ 78.6 77.6 92.2 92.2 88.3
BrowseComp 81.7 w/o Agentic Orchestration
82.8 🟒 w/ Agentic Orchestration		 75.5 67.9 67.6 91.2 πŸ₯‡ 76.2 90.4
DeepResearch-Bench2 43.7 w/o Agentic Orchestration
47.5 w/ Agentic Orchestration		 37.5 45.1 48.6 🟒 52.4 55.7 57.9 πŸ₯‡
XBench-DS-2510 92.0 πŸ₯‡ 86.0 71.0 80.0 85.0 84.0 90.0
βš™οΈ Engineering Tasks
MLE-Lite 65.2 🟒 43.9 37.9 39.4 72.7 72.7 77.3 πŸ₯‡
SciCode 44.3 🟒 44.3 36.6 39.5 58.7 59.0 πŸ₯‡ 56.1
SWE-bench Pro 61.2 🟒 43.6 49.5 59.6 63.3 64.6 64.6 πŸ₯‡
SWE-bench Multilingual 72.7 🟒 58.7 67.2 71.4 80.8 81.3 πŸ₯‡ 74.1
πŸ“‹ Instruction Following
IFBench 81.6 πŸ₯‡ 80.6 64.4 44.6 72.1 53.4 72.7
AA-LCR 81.0 🟒 74.2 71.7 73.2 88.7 πŸ₯‡ 79.7 84.0
πŸ§ͺ Scientific Research
FrontierScience-Olympiad 81.0 πŸ₯‡ 79.0 60.3 70.0 69.0 67.0 73.0
FrontierScience-Research 48.3 πŸ₯‡ 40.0 2.9 6.7 21.7 16.7 26.7
HiPhO 45.3 46.4 πŸ₯‡ 37.7 39.5 42.4 39.7 46.2
HLE w/ tools 55.1 🟒 47.6 36.2 33.4 56.0 62.5 64.5 πŸ₯‡
CMT 80.0 πŸ₯‡ 72.0 20 26.0 34.0 40.0 50.0
CritPt 20.0 🟒 4.3 0.3 2.9 23.4 19.1 32.3 πŸ₯‡
πŸ€– General Agentic Tasks
Agents' Last Exam (ALE-CLI) 19.1 🟒 10.2 3.81 14.3 27.6 28.5 28.6 πŸ₯‡
VitaBench 45.4 🟒 38.7 35.6 44.6 59.0 πŸ₯‡ 58.7 46.8
PinchBench 90.0 πŸ₯‡ 76.9 88.1 88.6 88.2 86.6 87.3
πŸ”¬ Scientific Agentic Tasks
ChemCOTBench 66.5 🟒 64.0 59.3 53.8 82.5 πŸ₯‡ 78.3 77.9
MatTools 42.6 47.1 🟒 15.9 38.6 24.3 53.2 πŸ₯‡ 45.6
OmniDocBench 91.5 πŸ₯‡ 89.5 89.9 89.6 91.1 - 85.8
> [!Note] > All benchmark results above are evaluated on the original full-precision (BF16) Sage-A1.5 weights. The Q8_0 quantization in this repository is near-lossless, with only a minimal deviation from the reference numbers. > > 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. ## Provided Files | File | Description | Size | | --- | --- | --- | | [`Sage-A1.5-Q8_0.gguf`](./Sage-A1.5-Q8_0.gguf) | Language model, Q8_0 quantization (MoE weights + built-in MTP layer) | 37.8 GB | | [`Sage-A1.5-mmproj.gguf`](./Sage-A1.5-mmproj.gguf) | Vision projector for multimodal (image) input | 0.9 GB | Notes: - The GGUF files use the `qwen35moe` architecture (hybrid attention + SSM MoE). **A recent build of your GGUF runtime with `qwen35moe` support is required** β€” older llama.cpp / Ollama / LM Studio versions will not load these files. - The multimodal projector is loaded automatically when it sits next to the main GGUF in runtimes that support mmproj auto-detection (e.g. LM Studio); otherwise pass it explicitly via `--mmproj` (llama.cpp). - For text-only deployments you can skip `Sage-A1.5-mmproj.gguf` entirely β€” the language model works standalone. - Memory footprint: ~35.2 GiB for the Q8_0 weights plus ~0.84 GiB for the projector, plus the KV cache which grows with the context length you allocate. A single 48 GB GPU comfortably serves the model at moderate context lengths; on smaller GPUs, split the layers across multiple cards (see `--tensor-split`) or lower `-ngl` to offload layers to CPU. ## Usage ### Download ```shell hf download InternScience/Sage-A1.5-Q8_0-GGUF # or, with the legacy CLI: # huggingface-cli download InternScience/Sage-A1.5-Q8_0-GGUF ``` ### llama.cpp [llama.cpp](https://github.com/ggml-org/llama.cpp) provides `llama-server` (an OpenAI-compatible API server) and `llama-cli` (interactive terminal). Use a recent version that supports the `qwen35moe` architecture. Install via package manager or build from source: ```shell # macOS / Linux via Homebrew brew install llama.cpp # or build from source git clone https://github.com/ggml-org/llama.cpp cd llama.cpp cmake -B build cmake --build build --config Release -j ``` The following command creates an OpenAI-compatible API endpoint at `http://localhost:8000/v1`: - **Standard Version** (multimodal, GPU offload): ```shell ./build/bin/llama-server \ -m Sage-A1.5-Q8_0.gguf \ --mmproj Sage-A1.5-mmproj.gguf \ --alias Sage-A1.5-Q8_0 \ -ngl 99 \ -fa on \ -c 65536 \ --port 8000 \ --jinja \ --reasoning-format deepthink ``` - **Text-Only** (skips the vision projector to free memory for a longer context): ```shell ./build/bin/llama-server \ -m Sage-A1.5-Q8_0.gguf \ --alias Sage-A1.5-Q8_0 \ -ngl 99 \ -fa on \ -c 131072 \ --port 8000 \ --jinja \ --reasoning-format deepthink ``` `--jinja` enables the model's built-in chat template with function-calling support; `--reasoning-format deepthink` keeps the model's self-review in `reasoning_content`, separate from the final answer in `content`. - **Quick test with `llama-cli`**: ```shell ./build/bin/llama-cli \ -m Sage-A1.5-Q8_0.gguf \ --mmproj Sage-A1.5-mmproj.gguf \ -ngl 99 -c 32768 \ --image /path/to/image.png \ -p "Describe this image in detail." ``` The model supports a context length of up to 262,144 tokens; allocate what you need with `-c`. The GGUF also carries the MTP (multi-token prediction) layer β€” self-speculative decoding with it depends on your llama.cpp version, so check the [llama.cpp documentation](https://github.com/ggml-org/llama.cpp/blob/master/docs) for the currently supported speculative-decoding flags. ### Ollama ```shell printf 'FROM ./Sage-A1.5-Q8_0.gguf\n' > Modelfile ollama create sage-a1.5-q8_0 -f Modelfile ollama run sage-a1.5-q8_0 ``` A recent Ollama version is required for the `qwen35moe` architecture, and vision (mmproj) import support varies by version β€” see the [Ollama documentation](https://github.com/ollama/ollama) for details. ### LM Studio Search for `Sage-A1.5 Q8_0` in LM Studio, or place both GGUF files under `~/.lmstudio/models/InternScience/Sage-A1.5-Q8_0-GGUF/`. LM Studio detects the vision projector automatically when the mmproj file sits next to the main GGUF, and image inputs can be attached directly in the chat view. ### 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](#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 β€” 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: ```python 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 β€” `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 β€” 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 β€” it is what tells the main model what its sub-agents can actually do. #### Verification Mode Verification Mode (see [Highlights](#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. Start `llama-server` with `--jinja` (as in the commands above) so the server renders the model's chat template and honors `chat_template_kwargs`. - **OpenAI Python client** (via `extra_body`): ```python from openai import OpenAI client = OpenAI(base_url="http://localhost:8000/v1", api_key="EMPTY") response = client.chat.completions.create( model="Sage-A1.5-Q8_0", 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**: ```shell curl http://localhost:8000/v1/chat/completions \ -H 'Content-Type: application/json' \ -d '{ "model": "Sage-A1.5-Q8_0", "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 β€” 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 When calling an OpenAI-compatible endpoint (e.g. `llama-server`), pass these fields in the request body; `repetition_penalty` is exposed as `repeat_penalty` in llama.cpp-based runtimes. ## 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](https://github.com/InternScience/Agents-A1/tree/main/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](https://huggingface.co/InternScience/Sage-A1.5) 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}, } ```