Apache-2.0multimodalRuns locally35B (card); ~36.0B (HF safetensors)

Sage-A1.5: Specs, License & How to Run

Shanghai AI Lab Apache-2.0 35B MoE science/agent VLM with orchestration and self-verification.

Official Sage-A1.5 benchmark overview figure from the InternScience model card — Official project image. Source · License: Apache-2.0.

License: Apache-2.0
Parameters: 35B (card); ~36.0B (HF safetensors)
HF likes: 223 (2026-10-23)
First seen: 2026-10-15

TL;DR. Sage-A1.5 is a multimodal model from Shanghai AI Laboratory (InternScience). Parameters: 35B (card); ~36.0B (HF safetensors).License: Apache-2.0. It is listed under ChatGPT, Closed LLM APIs.

Developer	Shanghai AI Laboratory (InternScience)
Type	multimodal
Architecture	Mixture-of-Experts VLM based on Qwen3.6-35B-A3B and Agents-A1. Hugging Face tags qwen3_5_moe; architecture Qwen3_5MoeForConditionalGeneration; pipeline_tag text-generation (API) / image-text-to-text (transformersInfo).
Parameters	35B (card); ~36.0B (HF safetensors)
Modalities	text, image
Context	256K native (card); recipes use 262,144; YaRN factor 2.0 → 512K
License	Apache-2.0 — Hugging Face cardData.license is apache-2.0 and Hub tags include license:apache-2.0 as of 2026-10-23. A raw fetch of resolve/main/LICENSE 404’d. Sibling FP8 and GGUF cards also tag apache-2.0. The related Agents-A1 GitHub repo is Apache-2.0; that is the prior A1 tree, not a LICENSE blob on this weight card. Not legal advice.
Hardware	Not yet verified — Card recipes name vLLM and SGLang at --max-model-len / --context-length 262144, including a 1-GPU vLLM example and a --language-model-only text-only path. HF usedStorage on the BF16 card is 71,928,317,664 bytes (~67.0 GiB); safetensors.parameters sum to 35,951,822,704. Q4_K_M GGUF usedStorage is 22,612,937,073 bytes. Peak VRAM is not independently measured. (not independently tested)
Hugging Face likes	223 (as of 2026-10-23)
Model card	https://huggingface.co/InternScience/Sage-A1.5
First seen	2026-10-15

Sage-A1.5 is a Mixture-of-Experts agentic VLM from Shanghai AI Laboratory (Hugging Face org InternScience). Official hosts we opened on 2026-10-23: the ungated card InternScience/Sage-A1.5 (created 2026-10-15T08:34:39Z; last modified 2026-10-22T14:08:48Z; 223 likes / 396 downloads), siblings Sage-A1.5-FP8 (12 likes / 116 downloads) and Sage-A1.5-Q4_K_M-GGUF (44 likes / 5,514 downloads; created 2026-10-21), the paper Sage-A1.5: From Long-Horizon Execution to Scientific Exploration with Capability Composition (submitted 2026-10-21), and the prior InternScience/Agents-A1 tree (578 stars; last push 2026-07-16). resolve/main/LICENSE on the weight card 404’d.

This page hangs off ChatGPT alternatives and closed LLM APIs. That mapping is editorial. We have no sourced quality comparison against ChatGPT or Ornith-1.5-122B. Head-to-head facts: Sage-A1.5 vs Ornith-1.5-122B.

The card says Sage-A1.5 is “a 35B Mixture-of-Experts agentic model based on Qwen3.6 and Agents-A1,” highlighting Agentic Orchestration, Verification, and Open Scientific Discovery. Those capability names stay vendor. Hugging Face tags include moe, vlm, vision, agentic, code, and tool-use.

What we verified (2026-10-23)

Field	Value	Source
Host (BF16)	huggingface.co/InternScience/Sage-A1.5 (not gated)	HF API
Created / last modified	2026-10-15T08:34:39Z / 2026-10-22T14:08:48Z	same
Likes / downloads	223 / 396	same
License tag	apache-2.0	HF `cardData`
LICENSE file fetch	resolve/main/LICENSE 404	this check
Pipeline tag	text-generation (API); image-text-to-text (transformersInfo)	HF API
Architecture tag	qwen3_5_moe	HF tags
Params (card)	35B MoE	card
safetensors.parameters	35,951,822,704 (~36.0B)	HF API
usedStorage (BF16)	71,928,317,664 bytes (~67.0 GiB)	same
Native context	256K (README); recipes 262,144	card
Base named	Qwen3.6-35B-A3B + Agents-A1	card
FP8 sibling	12 likes / 116 downloads	HF API
Q4_K_M GGUF	44 likes / 5,514 downloads; usedStorage 22,612,937,073	HF API
Agents-A1 GitHub	578 stars; last push 2026-07-16	GitHub API

The card name is 35B. Hugging Face’s parameter sum is ~36.0B. We keep both quotes.

How to run (from the card, not tested)

The card publishes vLLM and SGLang recipes with --reasoning-parser qwen3 and --tool-call-parser qwen3_coder. Standard vLLM (copied; 1 GPU, 262K):

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

A --language-model-only variant skips the vision encoder. MTP / EAGLE speculative flags are also on the card. forkalt has not run these commands.

Agentic Orchestration is opt-in: register a spawn_subagent tool. Verification Mode is opt-in: chat_template_kwargs.verification = true on a non-empty system message. Both default off.

For windows beyond 256K the card suggests YaRN (factor 2.0 ≈ 512K) and names the matching env flags. Enable it only when the workload needs it.

License

Card tag Apache-2.0. Same tag on the FP8 and Q4_K_M-GGUF siblings. A raw LICENSE fetch 404’d on 2026-10-23. The related Agents-A1 LICENSE is Apache-2.0 on GitHub; that file is the prior A1 repo, not this weight card. Not legal advice.

What we did not invent

VRAM, tokens/s, and “SOTA / narrows the gap with frontier models” stay vendor.
Card table rows (GAIA, BrowseComp, SWE-bench Pro, FrontierScience-*, …) stay self-reported. The table’s Ornith-1.5 column is upstream, not a forkalt ranking.
We do not rank Sage against ChatGPT or Ornith.

Related agent weights: Ornith-1.5-122B — Sage-A1.5 vs Ornith-1.5-122B — Flint (local coding MoE), Reflection Beam, Hy-Memory. Serving: vLLM.

Benchmarks

Quoted from the cited source. forkalt has not independently reproduced these numbers.

Name	Score	Source
GAIA (card table)	95.1 (self-reported)	https://huggingface.co/InternScience/Sage-A1.5
XBench-DS-2510 (card table)	92.0 (self-reported)	https://huggingface.co/InternScience/Sage-A1.5
SWE-bench Pro (card table)	61.2 (self-reported)	https://huggingface.co/InternScience/Sage-A1.5
FrontierScience-Research (card table)	48.3 (self-reported)	https://huggingface.co/InternScience/Sage-A1.5

Deploy options

local: Hugging Face weights (BF16)
local: Hugging Face weights (FP8)
local: Hugging Face GGUF (Q4_K_M)
local: vLLM serve recipe (card)
local: SGLang launch_server recipe (card)

Comparisons

Sage-A1.5 vs Ornith-1.5-122B

FAQ

Is Sage-A1.5 open weight and Apache-2.0?

The ungated card InternScience/Sage-A1.5 (created 2026-10-15T08:34:39Z; last modified 2026-10-22T14:08:48Z; 223 likes / 396 downloads in the 2026-10-23 API check) tags license apache-2.0. resolve/main/LICENSE 404’d. Sibling FP8 (12 likes) and Q4_K_M-GGUF (44 likes / 5,514 downloads) also tag apache-2.0. Not legal advice.

Is it a chatbot or a science agent?

The card describes a 35B MoE agentic model for long-horizon scientific problem solving, with three named capabilities: agentic orchestration, verification, and open scientific discovery. It is also tagged vlm / vision and ships an image-text chat template. Orchestration is opt-in via a spawn_subagent tool; verification is opt-in via chat_template_kwargs.verification. It is not a hosted search UI.

What is the context length?

The README says Sage-A1.5 natively supports a 256K context window. Serve recipes use 262,144 tokens. YaRN factor 2.0 is documented as a 512K override (SGLang SGLANG_ALLOW_OVERWRITE_LONGER_CONTEXT_LEN or vLLM VLLM_ALLOW_LONG_MAX_MODEL_LEN). Those are card figures, not a forkalt measurement.

Does it beat Ornith-1.5-122B or ChatGPT?

No sourced bake-off on this site. The card table includes a vendor Ornith-1.5 column and names Kimi-K3, GLM-5.3, and GPT-5.6-sol. Those rows stay self-reported. See Sage-A1.5 vs Ornith-1.5-122B for identification facts only.

Is github.com/InternScience/Agents-A1 the Sage-A1.5 weights repo?

No. That Apache-2.0 repo (578 stars; last push 2026-07-16 in the 2026-10-23 GitHub check) is the prior Agents-A1 project. Sage-A1.5 weights we opened live on the Hugging Face cards. The card’s GitHub link points at that A1 tree.

Sources & verification

Last verified 2026-10-23 · Last updated 2026-10-23· Edited by forkalt editors · Editorial policy

License notes are not legal advice. Read the original LICENSE. Trademark & disclaimer.