Sage S2.1 Flash is a scientific multimodal model from Shanghai Artificial Intelligence Laboratory, released on 2026-11-16. It uses a 321B-parameter MoE architecture and the Sage SciPerception unified representation to accept more than ten scientific input types, including molecular graphs, crystal structures, protein sequences and 3D structures. It is built for multi-step research work such as information retrieval, programming and operating professional scientific software, checking intermediate results against visual feedback. Weights are published under MIT in BF16 and FP8, and the model scored 90.3 on MLE-bench Lite.
A solid 321B-parameter MoE language model from Shanghai Artificial Intelligence Laboratory. A pragmatic middle-ground choice when you need open weights without a flagship-sized footprint. Newly released, so production-readiness is still being shaken out.
Generated from this model’s benchmarks and ranking signals. Editor reviews refine it over time.
Access model weights, configuration files, and documentation.
No benchmark data available for this model yet.
Benchmark data sourced from: Hugging Face, Arena.ai, Artificial Analysis.
The top devices for this model at 4-bit, ranked by fit and speed.
| Device | Grade | Speed | VRAM |
|---|---|---|---|
| NVIDIA Rubin GPUNVIDIA | SS | 718.5 tok/s | 197.2 GB |
| Rubin R100 UltraNVIDIA | SS | 89.8 tok/s | 197.2 GB |
| TPU 8i MaxGoogle | SS | 70.2 tok/s | 197.2 GB |
| AMD Instinct MI355XAMD | SS | 32.7 tok/s | 197.2 GB |
| HP ZGX Fury G1n AI Station |

Explore the Provider
Aggregate stats, leaderboard, release timeline, and benchmark coverage across every Shanghai Artificial Intelligence Laboratory model we track.
| SS |
| 29.0 tok/s |
| 197.2 GB |
See how different quantization levels affect VRAM requirements and quality for this model.
| Format | VRAM Required | Quality | |
|---|---|---|---|
| Q2_K | 129.8 GB | Low | |
| Q4_K_MRecommended | 197.2 GB | Good | |
| Q5_K_M | 229.3 GB | Very Good | |
| Q6_K | 267.8 GB | Excellent | |
| Q8_0 | 348.1 GB | Near Perfect | |
| FP16 | 653.0 GB | Full |
Sage S2.1 Flash is an open-weight, 321-billion-parameter mixture-of-experts (MoE) multimodal model developed by Shanghai Artificial Intelligence Laboratory. Released under the permissive MIT license, the model targets an underserved niche in local AI deployments: scientific workflows that demand autonomous reasoning, code execution, and visual verification. Rather than acting as a standard text-and-vision chatbot, Sage S2.1 Flash is engineered as an interactive research agent capable of driving scientific software, analyzing complex domain-specific data, and checking intermediate visual outputs before taking subsequent actions.
The model is released with both BF16 and FP8 weights on Hugging Face and ModelScope. In the broader ecosystem, it directly challenges proprietary and frontier research systems by proving that domain-specialized multimodal reasoning can be decoupled from closed APIs. By incorporating Shanghai AI Lab's Sage SciPerception architecture, the model accepts more than ten distinct scientific inputs, including molecular graphs, crystal structures, protein sequences, 3D biological models, and raw sensor spectra, while retaining general-purpose tool-calling and code-generation proficiency.
For engineers and lab researchers who need to run a local AI model 321B parameters in 2026 without data leaving their security perimeter, Sage S2.1 Flash represents one of the largest and most specialized open releases available. It bridges the gap between high-level agentic planning and low-level scientific computing, excelling at multi-turn workflows where an agent must write a script, execute an analysis package, inspect the generated plot or render, and correct errors on the fly.
Sage S2.1 Flash is built on a sparse mixture-of-experts (MoE) transformer backbone totaling 321B parameters. Although the exact count of active parameters per token is not publicly specified in the base technical documentation, the sparse routing design provides significant throughput benefits during inference. In an MoE architecture, only a fraction of the total network weights are activated for each token, allowing the compute footprint per step to stay significantly lower than a comparable 300B+ dense model.
The core technical differentiator is the Sage SciPerception unified representation framework:
Because the full 321B parameter set must remain in memory even when only a subset is activated per token, the memory footprint remains substantial.
Sage S2.1 Flash is purpose-built for verifiable agent execution. In real-world scientific computing, pure text models often produce plausible-sounding code that fails when encountering external software state changes. Sage S2.1 Flash counters this by relying on visual and numerical feedback loops across diverse scientific domains.
The model can interact with desktop and CLI-based scientific software, observing interface state through visual snapshots. Documented agent workflows include:
In coding and tool automation, the model is competitive with leading frontier systems. It achieved a top-ranking score of 90.3 on the MLE-bench Lite benchmark, demonstrating reliable proficiency in writing, debugging, and optimizing end-to-end machine learning pipelines. Sage S2.1 Flash for coding workflows functions as an autonomous engineer capable of refactoring scripts, interpreting stack traces, and validating outputs against unit tests.
The model achieved first-place benchmark scores across professional evaluation sets including OfficeQA Pro (72.2), Harvey LAB (25.0), and SpreadsheetBench 2 (37.1). This makes it well-suited for parsing dense technical literature, extracting experimental parameters from complex multi-column PDFs, and organizing tabular experimental outputs into structured datasets.
Operating a 321B-parameter model on private infrastructure requires careful capacity planning. Because all 321B parameters must reside in fast memory (VRAM or unified system memory) to avoid severe I/O bottlenecks during generation, memory capacity is the primary constraint.
Memory requirements depend heavily on your chosen quantization format. The figures below reflect the base weights plus a modest KV cache allocation:
When evaluating hardware requirements for running a model of this scale locally, consumer setups face physical limitations.
Can you run a 321B model on consumer GPUs? A single GPU cannot load this model under any practical quantization. Even an RTX 4090 with 24GB VRAM can only hold a fraction of the network. To run Sage S2.1 Flash locally on consumer hardware, you need multi-GPU pooling or unified memory:
Because of MoE routing efficiency, generation speeds are higher than a 321B dense model once the weights are resident in fast memory:
For local deployment, using an engine with robust MoE and vision projector support is critical. While Ollama provides the easiest command-line interface for local testing once community GGUF ports are published to the registry, production deployments should leverage specialized serving stacks:
1# Example deployment via vLLM for FP8 weights on an 8-GPU node2vllm serve internlm/Sage-S2.1-Flash-FP8 \3 --tensor-parallel-size 8 \4 --dtype fp8 \5 --trust-remote-code \6 --max-model-len 16384
For quantized workstation testing via llama.cpp:
1./llama-cli -m sage-s2.1-flash-Q4_K_M.gguf \2 --threads 16 \3 -ngl 99 \4 -c 8192 \5 -p "Analyze the following crystal structure parameters..."
Sage S2.1 Flash sits in an uncommon category: an open-weight, 300B+ MoE architecture tuned specifically for multimodal science, code, and agentic workflows.
DeepSeek's Flash variants are widely adopted for general programming, math, and conversational reasoning at low inference overhead. On standard machine learning engineering benchmarks, Sage S2.1 Flash holds a competitive edge, scoring 90.3 on MLE-bench Lite compared to DeepSeek-V4.1-Flash-1017 at 89.2. While DeepSeek models provide exceptional general coding throughput, Sage S2.1 Flash is tailored for cross-disciplinary research: it directly parses raw biomedical structures and domain instrumentation data that DeepSeek cannot natively ingest.
Against GLM-5.3-Flash, Sage S2.1 Flash demonstrates stronger execution in multi-step professional tool use and automated document handling, claiming first place on OfficeQA Pro and SpreadsheetBench 2. GLM models often excel in native bilingual conversational flow and general agent interfaces, but Sage S2.1 Flash is the preferred choice for laboratory practitioners who require the model to iteratively drive external software tools (such as Fiji, 3D Slicer, and Blender) using visual feedback to confirm results.
If your primary local workload consists of standard Python web services, general NLP, or simple vision-language tasks, lighter 32B to 70B dense models will be far easier to host. However, if your team requires an unconstrained, MIT-licensed foundation model to automate deep scientific workflows, parse specialized research modalities, and run verified machine learning engineering pipelines locally, Sage S2.1 Flash provides capabilities unmatched by standard general-purpose checkpoints.
Energy cost on Framework Desktop (AMD Ryzen AI Max+ PRO 495, 192GB) (~1.1 tok/s, Q4_K_M) vs flagship API pricing.
| Source | Cost per 1M tokens |
|---|---|
Local (energy only)Sage S2.1 Flash on Framework Desktop (AMD Ryzen AI Max+ PRO 495, 192GB) · ~1.1 tok/s · 120W | $3.59 |
GPT-6.1 TerraOpenAI · in $1.00 · out $5.00 | $2.20 |
Claude Fable 5.2Anthropic · in $8.00 · out $40.00 | $17.60 |
Gemini 4 HydrogenGoogle · in $1.50 · out $7.50 | $3.30 |
Grok 4.5xAI · in $2.00 · out $6.00 | $3.20 |
API prices blended at 70% input / 30% output.
Hardware amortisation not included. Run the full ROI calculator for payback math.
Cheapest current cloud rentals with at least 197 GB VRAM, refreshed hourly.
Advertising disclosure: we earn commissions when you shop through the links below.
| Option | Cost / GPU-hour |
|---|---|
AMD Instinct MI350XDigitalOcean · Spot · 288 GB VRAM | $3 |
AMD Instinct MI325XDigitalOcean · On-Demand · 256 GB VRAM | $3.8 |
AMD Instinct MI355XDigitalOcean · Spot · 288 GB VRAM | $4.5 |
AMD Instinct MI350XRunPod · Secure · 288 GB VRAM | $5.49 |
NVIDIA B300DigitalOcean · Spot · 288 GB VRAM | $6.17 |
Per-GPU rate across RunPod, the Vast.ai marketplace, DigitalOcean, and Vultr.
Spot tier is interruptible. Plan for restarts when comparing against on-demand prices.