Magma-8B

Microsoft, United States · February 2025

Microsoft Research

Foundation VLAOpen weights, MITAlso written Magma
Parameters
8.9 B
The Hugging Face safetensors index reports exactly 8,903,066,496 BF16 parameters for the released checkpoint.
GPU memory
17 GB
reported by the authors
Inference latency
1100 ms
Not stated.
Weights
MIT
downloadable checkpoint

What it is

Magma is a generalist agent model rather than a dedicated robot policy: the same 8.9B parameter checkpoint drives GUI navigation and robot manipulation, and its pretraining recipe adds Set-of-Mark and Trace-of-Mark annotations so that actionable objects and their motion are grounded in the image. On SimplerEnv it reports 52.3 percent on Google Robot and 35.4 percent on Bridge zero-shot, against 31.7 and 14.5 percent for OpenVLA. It is the only model in this comparison that publishes a real memory measurement, 17 GB peak in bfloat16 and 7 GB in 4-bit, which puts it inside a single 24 GB consumer card without tricks. The weak spot for a robotics buyer is timing and aggregation: the only published latency is about 1.1 s per inference with no GPU named, no action chunk or control frequency is documented, and the SimplerEnv figures average visual matching with variant aggregation, so they are not directly comparable to models that report visual matching alone. The model card restricts use to research and the repository has had no functional release since 2025.

Architecture

Backbone
LLaMA-3-8B language model with a CLIP ConvNeXt-XXLarge vision encoder (LAION trained)
Action head
Autoregressive token generation by the language model for both UI actions and 7-DoF robot gripper states. Pretraining adds Set-of-Mark labels for action grounding and Trace-of-Mark labels for motion planning. There is no separate diffusion or flow matching action expert.
Parameters
The Hugging Face safetensors index reports exactly 8,903,066,496 BF16 parameters for the released checkpoint. The paper's own comparison table lists the model size as 8.6B, and the Hugging Face model card text says 9B. The three figures disagree; the 8,903M value is the one that can be read off the actual weights.
Pretraining data
About 39M samples in total: 2.7M UI screenshots (SeeClick, Vision2UI), 9.4M image, language and action triplets from 970K Open X-Embodiment trajectories, more than 25M samples from about 4M instructional video clips (Ego4D, Epic-Kitchens, Something-Something v2), and 1.2M image text pairs for multimodal understanding (ShareGPT4V, LLaVA-1.5, OCR datasets).
Embodiments
Google Robot (SimplerEnv simulation), Bridge / WidowX (SimplerEnv simulation), WidowX 250 arm (real hardware), Desktop and mobile GUIs (UI navigation)

What hardware it needs

This is the section most people came for, so it is worth being precise about which numbers are measured and which are arithmetic.

Computed from 8.9 B parameters. Weights only, so treat it as a floor: activations, image encoder intermediates and the CUDA context come on top, in practice 30 to 50 percent more for inference.
PrecisionWeightsNote
fp3233 GBTraining master weights, rarely used for inference
bf16 or fp1617 GBThe usual way these checkpoints are served
int88.3 GBQuantised, expect some loss of precision on fine motions
int44.1 GBAggressive quantisation, verify success rate before trusting it

Smallest real card that fits the bf16 weights plus a 40 percent runtime allowance: 24 GB. That is an estimate for inference, not for fine tuning. Full fine tuning also holds optimiser state and gradients, which typically costs several times the weights unless you use LoRA or a comparable adapter.

Published figures. These come from the model authors, not from this site.
Inference memory, reported17 GB
Inference latency, reported1100 ms on Not stated. The GitHub README reports 1.1 s inference time for both the bfloat16 and the 4-bit configuration but does not name the GPU, the input resolution or the output length, so this number cannot be transferred to other hardware. source
Checkpoint download17.82 GB

The README reports 17 GB peak memory in bfloat16 and 7 GB in 4-bit, so a single 24 GB card is sufficient in bfloat16 and a 12 GB card should suffice quantized. The GPU used for the measurement is not named. No fine-tuning VRAM figure is published; pretraining used H100 and MI300 GPUs.

Reported results

Grouped by what the evaluation actually asked. Values from different suites are not comparable with each other, so each one keeps its suite and split.

Simulation

Reproducible benchmark suites. Everyone runs the same episodes, so the numbers can be compared inside a suite.

  • SimplerEnv Google Robot, averaged across visual matching and variant aggregation, zero-shot
    52.3%
    success rateThe table caption states results are averaged across visual matching and variant aggregation, so this is not comparable to visual matching only numbers. Baselines in the same table: OpenVLA (8B) 31.7, RT-1-X (35M) 34.2, Octo (93M) 6.0.source
  • SimplerEnv Bridge / WidowX, averaged across visual matching and variant aggregation, zero-shot
    35.4%
    success rateSame averaging caveat. Baselines: OpenVLA 14.5, Octo 15.9, RT-1-X 1.1.source

Real world

Rollouts on physical hardware. The setups differ, so read these as evidence, not as a ranking.

  • AITW fine-tuned, overall
    67.3%
    success rateUI navigation, not robotics.source
  • Mind2Web fine-tuned
    57.2%
    element accuracyUI navigation, not robotics.source
  • Mind2Web fine-tuned
    45.4%
    step success rateUI navigation, not robotics.source

Fine tuned tasks

No results in this category are published for this model.

On real hardware

Evaluated on a WidowX 250 arm across four manipulation tasks with 50 demonstration trajectories each. The paper reports that Magma substantially outperforms the baseline and that OpenVLA can hardly accomplish tasks such as picking and placing a hotdog sausage or putting a mushroom in a pot, but it does not publish per task success percentages for the real arm.

Fine tuning it yourself

Training code and the pretraining datasets (Magma-820K, Magma-Video-ToM, Magma-OXE-ToM) are published on Hugging Face. The pretraining configuration is batch size 1024, learning rate 1e-5, maximum sequence length 4096 and 3 epochs on H100 and MI300 GPUs. LIBERO few-shot fine-tuning was demonstrated with 10 trajectories per task, but the paper presents that result as a figure rather than a numeric table.

Where it helps, where it does not

Strengths

  • Real published memory numbers, 17 GB in bfloat16 and 7 GB in 4-bit, which is rare among VLAs
  • One checkpoint covers GUI agents and robot manipulation, useful if both surfaces are in scope
  • MIT license on code and weights, with datasets and demo spaces published as well
  • Set-of-Mark and Trace-of-Mark supervision gives explicit, inspectable action grounding rather than opaque action tokens

Limits

  • About 1.1 s per inference and no documented action chunking, which makes closed loop control at any meaningful rate the open question this model does not answer.
  • The SimplerEnv figures average visual matching and variant aggregation, so putting 52.3 next to a visual matching only number from another model is a false comparison.
  • Real robot evidence is limited to a WidowX 250 arm with 50 trajectories per task. No bimanual and no dexterous hand results.
  • The model card designates the model for research use only and states it is not designed for production or high risk scenarios.
  • No control frequency, no action chunk size and no fine-tuning VRAM figure are published.
  • No successor. The last functional repository change dates to 2025 and no Magma-2 exists as of August 2026.

Sources

Everything on this page was taken from these documents. Where they disagree with what you read here, they win.

Entry last checked 2026-08-11.

Try a policy on a real arm

A physical SO-100 is online and free to drive in the browser, and five of the models in this arena can be fine tuned on a dataset you record with your own.