InternScienceAgents-A1Qwen3_5MoeForConditionalGeneration

Agents A1 — Hardware Requirements & GPU Compatibility

ChatVisionFunctions

Agents-A1 is InternScience's 35-billion-parameter mixture-of-experts agentic model, built to handle long-horizon tool use, engineering, scientific research, and instruction-following tasks rather than general chit-chat. It routes to roughly 3.9 billion active parameters per token and is trained with a three-stage recipe: broad supervised fine-tuning on agentic behaviors, domain-specialist teacher models, and multi-teacher multi-domain on-policy distillation that transfers that expertise back into one general model. The card reports it approaching or matching frontier-scale systems like GPT-5.5 and Kimi-K2.6 on agentic benchmarks such as BrowseComp, GAIA, and SciCode despite its far smaller size, and it needs a capable multi-GPU workstation to run at full precision, less once quantized. Context length is 262,144 tokens. It is released under the Apache 2.0 license, permitting unrestricted commercial and research use, and was published in June 2026; the same family later added a smaller 4B model for local deployment.

35.1B params · MoE3.9B active262K contextApache 2.0Jun 202610.4K downloads653 likes6.5K quant downloads

Your hardware

Verdicts below are for the reference RTX 4090. Set yours to personalise the page.

Detection reads your browser's GPU name with no data sent anywhere.

Runs well on a 24 GB GPU

Runs from 12 GB (IQ2_S · 64K)Best on 24 GB (Q4_K_M · ~171 tok/s · 128K)Macs from 16 GB (IQ2_M · ~57 tok/s)

Apache 2.0

Reference GPU · RTX 4090 24 GB

Estimates, ±20% · how we estimate

Run Agents A1

One recommended way. Every other method, OS and quant is in the step-by-step guide.

Other ways

each opens step-by-step

Run in cloud

Fits on RTX A6000 (48 GB) (26 GB headroom) · Q4_K_M

Generation speed: ~143 tok/s; generation speed
Cost per 1M output tokens: $0.64; per 1M output tokens

Deployon RunPodBest optionfrom $0.33/hr

Compare GPUs →

Run Agents A1

21.5 GB VRAM at 16K context · fits the RTX 4090 (2.5 GB free)

Works on macOS, Windows and Linux.

Install Ollama (Linux)
$curl -fsSL https://ollama.com/install.sh | sh
Install Ollama (macOS / Windows)
Download and run the app; it installs the ollama command.
Download Ollama ↗
Run Q4_K_M
$ollama run hf.co/mradermacher/Agents-A1-i1-GGUF:Q4_K_M
Raise the context window
Type it inside the chat. For the API server instead, start it with OLLAMA_CONTEXT_LENGTH=16384 ollama serve.
$/set parameter num_ctx 16384
A larger context needs more memory than the weights alone.
Use the API
Ollama serves its API on http://localhost:11434.

Works on macOS, Windows and Linux.

Install LM Studio
Desktop app for macOS (Apple silicon), Windows and Linux.
lmstudio.ai ↗
Download the model
In the app press Ctrl/Cmd + Shift + M and paste https://huggingface.co/mradermacher/Agents-A1-i1-GGUF, then pick Q4_K_M. Or use the CLI (run LM Studio once first):
$lms get https://huggingface.co/mradermacher/Agents-A1-i1-GGUF@Q4_K_M
Open on Hugging Face ↗
Load and chat
Open the model loader, select the model and set Context Length to 16384. Leave GPU offload at maximum if it fits.
Start the local server (optional)
$lms server start

Works on macOS, Windows and Linux.

Install llama.cpp (macOS)
MacPorts, Nix and conda-forge also work.
$brew install llama.cpp
Install options ↗
Install llama.cpp (Windows)
$winget install llama.cpp
Install llama.cpp (Linux)
Download the prebuilt llama-*-bin-ubuntu-* archive for your GPU backend (CPU, Vulkan) from the releases page, unpack it and run the binaries from that folder. Homebrew (brew install llama.cpp) and conda-forge also work on Linux.
llama.cpp releases ↗
Start the server
Downloads the model on first run. The chat UI and an OpenAI-compatible API are at http://localhost:8080.
$llama-server -hf mradermacher/Agents-A1-i1-GGUF:Q4_K_M -c 16384 -ngl 99 --port 8080

Works on macOS, Windows and Linux.

Option A: Docker Model Runner
Docker Desktop: Settings > AI > Enable Docker Model Runner. Docker Engine on Linux: install the docker-model-plugin package.
$docker model run hf.co/mradermacher/Agents-A1-i1-GGUF:Q4_K_M
Model Runner docs ↗
The quant tag after the repo is how Hugging Face cards write it; if your Docker version rejects it, drop the tag and pick the quant by repo.
Option B: llama.cpp server image (NVIDIA)
Needs the NVIDIA Container Toolkit. Models are cached in the llama-cache volume. Open http://localhost:8080.
$docker run --gpus all -p 8080:8080 -v llama-cache:/root/.cache ghcr.io/ggml-org/llama.cpp:server-cuda -hf mradermacher/Agents-A1-i1-GGUF:Q4_K_M -c 16384 -ngl 99 --host 0.0.0.0 --port 8080
Image tags (cuda13, vulkan, rocm) ↗
Cache path inside the image is not in the docs; remove the -v flag if the volume complains.

Linux only.

Install vLLM
Linux only; on Windows run it inside WSL2. NVIDIA GPU shown; AMD ROCm, Intel and CPU builds have their own install pages.
$uv venv --python 3.12 --seed
$source .venv/bin/activate
$uv pip install vllm --torch-backend=auto
vLLM installation ↗
Serve the model
OpenAI-compatible API at http://localhost:8000/v1.
$vllm serve InternScience/Agents-A1 --max-model-len 16384
Runs the original weights, not the Q4_K_M GGUF; needs ~77 GB of GPU memory.

macOS only.

MLX supports only macOS, so follow these steps for that setup.

Install mlx-lm
$pip install mlx-lm
Chat in the terminal
$mlx_lm.chat --model mlx-community/Agents-A1-4bit
Downloads the model on first run. The MLX repo has its own quantisation; it may not match the quant selected above.
Or start a local server
Listens on http://localhost:8080.
$mlx_lm.server --model mlx-community/Agents-A1-4bit

Works on macOS, Windows and Linux.

Install Jan
jan.ai ↗
Find the model
Open the Hub tab, paste this model ID in Search, and choose Q4_K_M if offered.
$mradermacher/Agents-A1-i1-GGUF
Use the local API (optional)
Settings > Local API Server > Start Server. It listens on http://127.0.0.1:1337.

Rent GPUs hourly to run Q4_K_M right there. Pricing comes from partner marketplaces and varies.

Available GPU rentals for Q4_K_M, prioritized by value
GPU	Speed	$/1M tokens	Rent
RTX A6000 (48 GB)CheapestCheapest26 GB free	~143 tok/s	$0.64	Deploy on RunPodfrom $0.33/hr (opens in new tab)	48 GB · 26 GB free · ~143 tok/s · $0.64 / 1M tokens
RTX PRO 4500 (32 GB)10 GB free	~159 tok/s	$0.59	Deploy on RunPodfrom $0.34/hr (opens in new tab)	32 GB · 10 GB free · ~159 tok/s · $0.59 / 1M tokens
A40 (48 GB)26 GB free	~134 tok/s	$0.73	Deploy on RunPodfrom $0.35/hr (opens in new tab)	48 GB · 26 GB free · ~134 tok/s · $0.73 / 1M tokens
A100 40GB18 GB free	~219 tok/s	$0.61	Deploy on Vast.aifrom $0.48/hr (opens in new tab)	18 GB free · ~219 tok/s · $0.61 / 1M tokens
RTX 5000 Ada (32 GB)10 GB free	~117 tok/s	$1.17	Deploy on RunPodfrom $0.49/hr (opens in new tab)	32 GB · 10 GB free · ~117 tok/s · $1.17 / 1M tokens
RTX 6000 Ada (48 GB)26 GB free	~166 tok/s	$1.00	Deploy on Vast.aifrom $0.60/hr (opens in new tab)	48 GB · 26 GB free · ~166 tok/s · $1.00 / 1M tokens
L40S (48 GB)26 GB free	~155 tok/s	$1.34	Deploy on Vast.aifrom $0.75/hr (opens in new tab)	48 GB · 26 GB free · ~155 tok/s · $1.34 / 1M tokens
A100 80GB58 GB free	~250 tok/s	$1.48	Deploy on Vast.aifrom $1.33/hr (opens in new tab)	58 GB free · ~250 tok/s · $1.48 / 1M tokens

Compare GPUs →

How Much VRAM Does Agents A1 Need?

Select a quantization level; the listed hardware adheres to it.

VRAM	+32K	+128K	File	Quality	On RTX 4090 (reference)
10.0 GB	10.3 GB	11.3 GB	9.5 GB	2-bit · i-quant extreme compression — Importance-weighted 2-bit, extreme compression — significant quality loss	Fits: 256K context · ~256 tok/s
10.9 GB	11.2 GB	12.2 GB	10.5 GB	2-bit · i-quant extra small — Importance-weighted 2-bit, extra small	Fits: 256K context · ~247 tok/s
11.3 GB	11.6 GB	12.6 GB	10.7 GB	2-bit · i-quant small — Importance-weighted 2-bit, small	Fits: 256K context · ~242 tok/s
12.2 GB	12.5 GB	13.5 GB	11.7 GB	2-bit · i-quant medium — Importance-weighted 2-bit, medium	Fits: 256K context · ~234 tok/s
14.0 GB	14.2 GB	15.3 GB	13.6 GB	3-bit · i-quant — Importance-weighted 3-bit	Fits: 256K context · ~219 tok/s
14.4 GB	14.7 GB	15.7 GB	12.2 GB	2-bit · small K-quant — 2-bit small K-quant	Fits: 256K context · ~215 tok/s
14.9 GB	15.1 GB	16.1 GB	14.5 GB	3-bit · i-quant extra small — Importance-weighted 3-bit, extra small	Fits: 256K context · ~212 tok/s
15.3 GB	15.6 GB	16.6 GB	15.3 GB	3-bit · i-quant small — Importance-weighted 3-bit, small	Fits: 256K context · ~209 tok/s
15.3 GB	15.6 GB	16.6 GB	12.9 GB	2-bit · K-quant — 2-bit quantization with K-quant improvements	Fits: 256K context · ~209 tok/s
15.7 GB	16 GB	17 GB	15.2 GB	3-bit small — 3-bit small quantization	Fits: 256K context · ~205 tok/s
16.2 GB	16.4 GB	17.4 GB	15.4 GB	3-bit · i-quant medium — Importance-weighted 3-bit, medium	Fits: 256K context · ~202 tok/s
17.5 GB	17.8 GB	18.8 GB	16.8 GB	3-bit medium — 3-bit medium quantization	Fits: 256K context · ~194 tok/s
17.9 GB	18.2 GB	19.2 GB	19.8 GB	4-bit · legacy — 4-bit legacy quantization	Fits: 256K context · ~191 tok/s
18.4 GB	18.6 GB	19.6 GB	18.1 GB	3-bit large — 3-bit large quantization	Fits: 256K context · ~188 tok/s
19.3 GB	19.5 GB	20.5 GB	18.7 GB	4-bit · i-quant compact — Importance-weighted 4-bit, compact	Fits: 256K context · ~183 tok/s
20.1 GB	20.4 GB	21.4 GB	19.9 GB	4-bit small — 4-bit small quantization	Fits: 256K context · ~178 tok/s
20.1 GB	20.4 GB	21.4 GB	21.8 GB	4-bit · legacy+ — 4-bit legacy quantization with offset	Fits: 256K context · ~178 tok/s
21.4 GB	21.7 GB	22.7 GB	21.2 GB	4-bit medium — 4-bit medium quantization — most popular sweet spot	Fits: 128K context · ~171 tok/s
24.5 GB	24.8 GB	25.8 GB	24 GB	5-bit small — 5-bit small quantization	Does not fit: needs 24.6 GB
25.4 GB	25.7 GB	26.7 GB	24.7 GB	5-bit medium — 5-bit medium quantization — good quality/size tradeoff	Does not fit: needs 25.5 GB
29.4 GB	29.6 GB	30.6 GB	28.5 GB	6-bit, very good — 6-bit quantization, very good quality	Does not fit: needs 29.4 GB
35.5 GB	35.7 GB	36.8 GB	36.9 GB	8-bit, near-lossless — 8-bit quantization, near-lossless	Does not fit: needs 35.6 GB
70.6 GB	70.9 GB	71.9 GB	~70.2 GB	16-bit · bfloat — Brain floating point 16 — preferred for training	Does not fit: needs 70.7 GB

Selected Q4_K_M

4-bit medium quantization

21.5GBat 16K context

Weights: 21.1 GB
KV cache: 0.2 GB
Overhead: 0.3 GB

Fits RTX 4090 (reference) · 2.5 GB free · ~171 tok/s

≥50 tok/s needs ≥206 GB/s · RTX 5090 ~236, RTX 3090 Ti ~171

Or rent RTX A6000 from $0.33/hr.

Which Hardware Can Run Agents A1?

Q4_K_M · 21.4 GB

Agents A1 (Q4_K_M) requires 21.4 GB of VRAM to load the model weights. For comfortable inference with headroom for KV cache and system overhead, 28+ GB is recommended. Using the full 262K context window can add up to 2.6 GB, bringing total usage to 24.1 GB. 7 GPUs can run it, including NVIDIA GeForce RTX 5090. 100 devices with unified memory can run Agents A1, including Mac Studio (M3 Ultra, 256GB), Mac Studio (M3 Ultra, 512GB), Mac Studio M1 Max (32 GB).

By VRAM tier · typical GPUs and Macs at Q4_K_M

Typical hardwareQuantVRAMtok/sContext that fits

8 GB · RTX 4060Doesn't fit — smallest quant needs 10.1 GB8 GB tier →

12 GB · RTX 3060 12GBIQ2_S · 11.4 GB · ~132 tok/s · 64K11.4~13264K12 GB tier →

16 GB · RTX 5060 Ti 16GBQ3_K_S · 15.8 GB · ~122 tok/s · 32K15.8~12232K16 GB tier →

24 GB · RTX 4090referenceQ4_K_M · 21.5 GB · ~171 tok/s · 128K21.5~171128K24 GB tier →

Mac 32 GB · Mac Mini M4Q4_K_M · 21.5 GB · ~28 tok/s · 256K21.5~28256KMacs →

Mac 64 GB · Mac Studio M4 MaxQ4_K_M · 21.5 GB · ~70 tok/s · 256K21.5~70256KMacs →

Alternatives to Agents A1

Similar models

6tok/s on RTX 4090

Model	Score	VRAM	tok/s	Why it is similar
Qwen3.6 35B A3BAlibaba	106.8	22 GB	~151	same size
MiniCPM V 4.7 35B A3BOpenBMB	—	22.1 GB	~169	same size
Llm Jp 4.1 33B Thinkingllm-jp	—	22.5 GB	~35	same size
Agnes 3.0 FlashAgnes-AI	—	22.1 GB	~35	same size
Ornith 1.5 35B A3Bornith-ai	—	22 GB	~151	same size
Intern S2 PreviewInternLM	—	73.2 GB	—needs > 24 GB	same size

Specifications

Publisher: InternScience
Family: Agents-A1
Parameters: 35.1B
Active parameters: 3.9B per token
Architecture: Qwen3_5MoeForConditionalGeneration
Context Length: 262,144 tokens
Vocabulary Size: 248,320
Release Date: 2026-06-22
License: Apache 2.0

Resources

Hugging Face model card (opens in a new tab)

Frequently Asked Questions

How much VRAM does Agents A1 need?

Agents A1 requires 21.4 GB of VRAM at Q4_K_M, or 70.6 GB at BF16. Full 262K context adds up to 2.6 GB (24.1 GB total).

VRAM = Weights + KV Cache + Overhead

Weights = 35.1B × 4.8 bits ÷ 8 = 21.1 GB

KV Cache + Overhead ≈ 0.3 GB (at 2K context + ~0.3 GB framework)

Fit ratings and hardware model lists check this model with room for a 16K-token context, which needs a little more memory.

KV Cache + Overhead ≈ 2.9 GB (at full 262K context)

VRAM usage by quantization

Q4_K_M

21.4 GB

Q4_K_M + full context

24.1 GB

Learn more about VRAM estimation →

Can NVIDIA GeForce RTX 4090 run Agents A1?

Yes, at Q4_K_M (21.4 GB) or lower. Higher quantizations like Q5_K_S (24.5 GB) exceed the NVIDIA GeForce RTX 4090's 24 GB.

What's the best quantization for Agents A1?

For Agents A1, Q4_K_M (21.4 GB) offers the best balance of quality and VRAM usage. Q5_K_S (24.5 GB) provides better quality if you have the VRAM. The smallest option is IQ2_XXS at 10.0 GB.

VRAM requirement by quantization

IQ2_XXS

10.0 GB~53%

IQ3_XS

14.9 GB~73%

Q3_K_M

17.5 GB~83%

Q4_K_M ★

21.4 GB~89%

Q5_K_S

24.5 GB~92%

BF16

70.6 GB~100%

★ Recommended — best balance of quality and VRAM usage.

Learn more about quantization →

Can I run Agents A1 on a Mac?

Yes — Mac Mini M2 Pro (16 GB) and 133 other Macs can run Agents A1, the smallest at IQ2_M. Apple Silicon uses unified memory, so the model shares RAM with the system. At Q4_K_M you need at least 21.4 GB of usable unified memory (RAM minus macOS overhead), from Mac Studio M1 Max (32 GB).

Can I run Agents A1 locally?

Yes — Agents A1 can run locally on consumer hardware. Popular tools include Ollama, LM Studio, and llama.cpp.

How fast is Agents A1?

At Q4_K_M, Agents A1 can reach ~118 tok/s on AMD Instinct MI350X. On NVIDIA GeForce RTX 4090: ~171 tok/s. Speed depends mainly on GPU memory bandwidth. Real-world results typically within ±20%.

tok/s = 1000 ÷ (active GB ÷ (bandwidth GB/s × efficiency) × 1000 + layers × routing ms)

Mixture-of-Experts: just the active experts get read for each token, along with a fixed routing cost per layer.

Example: NVIDIA B200 → 2.4 GB active ÷ (8000 × 0.65) = 0.46 ms, plus 40 layers × 0.055 ms = 2.20 ms, so 1000 ÷ 2.66 ms = ~376 tok/s

Estimated speed at Q4_K_M (21.4 GB)

NVIDIA B200

~376 tok/s

NVIDIA GeForce RTX 4090

~171 tok/s

NVIDIA B300

~376 tok/s

NVIDIA GH200 Grace Hopper Superchip

~339 tok/s

Real-world results typically within ±20%. Speed depends on batch size, quantization kernel, and software stack.

Learn more about tok/s estimation →

What's the download size of Agents A1?

At Q4_K_M, the download is about 21.06 GB. The full-precision BF16 version is 70.21 GB. The smallest option (IQ2_XXS) is 9.65 GB.

Which GPUs can run Agents A1?

7 consumer GPUs can run Agents A1 at Q4_K_M. Top options include AMD Radeon RX 7900 XTX, NVIDIA GeForce RTX 3090.

Which devices can run Agents A1?

100 devices with unified memory can run Agents A1 at Q4_K_M (21.4 GB), including AMD Ryzen AI 9 HX 370 (Strix Point) Laptop, ASUS Ascent GX10, Beelink GTR9 Pro (Ryzen AI Max+ 395, 128 GB), Framework Desktop (Ryzen AI Max+ 395, 128 GB). Apple Silicon Macs use unified memory shared between CPU and GPU, making them well-suited for local LLM inference.