Mirror of meta-models/Muse-Glimmer-30B-ExecuTorch-PTE. All credit to the original authors. Not an official release from them. License unchanged (apache-2.0); see LICENSE, USAGE_POLICY.md.

Muse Glimmer 30B — ExecuTorch PTE

Authors: Meta Superintelligence Lab
Model Release Date: August 2026
License: Apache 2.0

[!NOTE] A PTE is the serialized artifact ExecuTorch produces for a PyTorch model, lowered and optimized for a specific target backend. This repo contains pre-exported PTEs for NVIDIA CUDA (sm80+ptx) and Apple Silicon (metal), for text-only and text+image, with and without DFlash speculative decoding. Everything you need to download and serve one is on this page. For export from source, custom quantization recipes, and runtime internals, see Muse Glimmer at ExecuTorch.

Muse Glimmer is a 30-billion-parameter causal language model with a dedicated perception encoder, distilled from Muse Spark and purpose-built for autonomous agentic tasks on consumer hardware. The model integrates multi-step reasoning, reliable tool use, multimodal understanding, and failure recovery into a single model that runs locally without requiring cloud infrastructure or network access.

Why an export, rather than a per-backend port. Local runtimes usually reimplement a model by hand for each target they support. That holds up for a plain text transformer; it does not hold up for a novel architecture with multimodal input and block-diffusion speculative decoding, where every backend would need its own rewrite of all three. ExecuTorch inverts that: the model and its decoding strategy are written once in PyTorch, and torch.export lowers the whole graph ahead of time — Triton on CUDA, MLX-native and custom Metal on Apple Silicon. The files in this repo are the output of that export. The PyTorch team's announcement makes the full argument.

---

⚠️ Read this before you download

This repo is 372 GB in total. A bare hf download meta-models/Muse-Glimmer-30B-ExecuTorch-PTE will try to pull all 16 variants. Always use --include to fetch exactly one variant (17.9–31.5 GB) plus the shared root files. The commands below do this.

On CUDA, the .pte is not the model. For every sm80+ptx variant the weights live in the .ptd (19–31 GB) and the .pte is only 16–35 MB. Both files are required. If you grabbed a 25 MB .pte and nothing works, this is why.

---

What's in this repo

16 variant directories, named by a fixed scheme:

muse-glimmer-<quant>-128K-<modality>-<decoding>-<backend>

What each directory contains

Every directory follows the same three rules — there are no exceptions:

1. Always: <dirname>.pte
2. If and only if sm80+ptx: <dirname>.ptd — the CUDA delegate blob holding the weights
3. If and only if text-image: pos_embed.bin — precomputed image position embeddings

A dflash directory is not an exception to rule 1. Target and drafter are exported together into that single .pte and share their token embeddings and output head, so the drafter costs far less than a second model would.

The artifacts are named after their own directory.

The artifacts are named after their own directory. They are not model.pte / aoti_cuda_blob.ptd — those are the filenames a local export_solo run writes, and they do not exist anywhere in this repo. Copy-pasting a quickstart that references them will fail. Use the $VARIANT pattern below and you cannot get this wrong.

Shared files at the repo root

tokenizer.json, tokenizer_config.json, chat_template.jinja, LICENSE, USAGE_POLICY.md. You need the three tokenizer/template files alongside whichever variant you pick — the server's --hf-tokenizer loads that directory to render prompts and tool definitions.

Download sizes (per directory, .pte + .ptd + pos_embed.bin)

---

Pick a variant

1. Backend — your hardware decides. Apple Silicon → metal. NVIDIA SM80+ → sm80+ptx. There is no CPU variant.
2. Quant — k-quant-17G for 24 GB of VRAM/unified memory, k-quant-dynamic for 32 GB. dynamic is measurably closer to full precision (see Fitting the Model on Your Device).
3. Modality — take text-image only if you actually send images; it costs memory and download size.
4. Decoding — dflash is meaningfully faster on capable GPUs but adds the drafter's memory and download cost. solo is the smaller, simpler starting point.

If you're unsure, start with muse-glimmer-k-quant-17G-128K-text-solo-<your backend>.

---

Download one variant

Set the variant once; every command below derives its paths from it.

REPO=meta-models/Muse-Glimmer-30B-ExecuTorch-PTE
VARIANT=muse-glimmer-k-quant-17G-128K-text-solo-sm80+ptx   # ← change this
LOCAL_DIR=./muse-glimmer-pte

hf download "$REPO" \
  --include "$VARIANT/*" \
  --include "tokenizer.json" \
  --include "tokenizer_config.json" \
  --include "chat_template.jinja" \
  --local-dir "$LOCAL_DIR"

This pulls one variant directory plus the shared tokenizer/template files. Anything without --include pulls all 372 GB.

Two things about hf download worth knowing here:

• --include takes one pattern per flag. Repeat it, as above.
• Never mix positional filenames with --include. hf download "$REPO" --include "$VARIANT/*" tokenizer.json silently warns Ignoring --include since filenames have being explicitly set and downloads only the tokenizer.

Add --dry-run to see the exact file list and total size before committing to the transfer:

hf download "$REPO" --include "$VARIANT/*" --dry-run

Resulting layout:

muse-glimmer-pte/
├── tokenizer.json
├── tokenizer_config.json
├── chat_template.jinja
└── muse-glimmer-k-quant-17G-128K-text-solo-sm80+ptx/
    ├── muse-glimmer-k-quant-17G-128K-text-solo-sm80+ptx.pte   #  16 MB — graph only
    └── muse-glimmer-k-quant-17G-128K-text-solo-sm80+ptx.ptd   #  20 GB — the weights

---

Build the runner

A download is not enough. The pre-exported PTEs let you skip the export step — they do not remove the need for the native runtime. The server launches a muse_glimmer_worker binary, and that must be built from source.

Set up an ExecuTorch checkout per the ExecuTorch README and build-from-source guide, then build the model-specific workflow preset from the repo root:

# NVIDIA CUDA (Linux or Windows)
(cd examples/models/muse-glimmer && cmake --workflow --preset muse-glimmer-cuda)

# Apple Silicon (macOS)
(cd examples/models/muse-glimmer && cmake --workflow --preset muse-glimmer-mlx)

Binaries land in cmake-out/examples/models/muse-glimmer/: solo_runner, dflash_runner, and muse_glimmer_worker. The server needs muse_glimmer_worker.

Install the server dependencies:

pip install -r examples/llm_server/python/requirements.txt

---

Serve

An OpenAI-compatible server. Run from the ExecuTorch repository root.

CUDA (sm80+ptx) — both --model-path and --data-path

VARIANT=muse-glimmer-k-quant-17G-128K-text-solo-sm80+ptx
LOCAL_DIR=./muse-glimmer-pte

python -m executorch.examples.models.muse_glimmer.serving.serve \
  --model-path "$LOCAL_DIR/$VARIANT/$VARIANT.pte" \
  --data-path  "$LOCAL_DIR/$VARIANT/$VARIANT.ptd" \
  --tokenizer-path "$LOCAL_DIR/tokenizer.json" \
  --hf-tokenizer "$LOCAL_DIR" \
  --worker-bin cmake-out/examples/models/muse-glimmer/muse_glimmer_worker \
  --model-id muse-glimmer-30B \
  --tool-parser atem \
  --host 127.0.0.1 --port 8000

Metal — self-contained .pte, no --data-path

VARIANT=muse-glimmer-k-quant-17G-128K-text-solo-metal
LOCAL_DIR=./muse-glimmer-pte

python -m executorch.examples.models.muse_glimmer.serving.serve \
  --model-path "$LOCAL_DIR/$VARIANT/$VARIANT.pte" \
  --tokenizer-path "$LOCAL_DIR/tokenizer.json" \
  --hf-tokenizer "$LOCAL_DIR" \
  --worker-bin cmake-out/examples/models/muse-glimmer/muse_glimmer_worker \
  --model-id muse-glimmer-30B \
  --tool-parser atem \
  --host 127.0.0.1 --port 8000

Adjustments per variant

• text-image — add --pos-embed-path "$LOCAL_DIR/$VARIANT/pos_embed.bin".
• dflash — no extra flag. --artifact-mode defaults to auto and detects the exported method contract. The block dimension is exported dynamically, so --dflash-block-length is selectable at runtime — but the exported range differs by backend: metal accepts [2, 16], sm80+ptx only [2, 4], and on sm80+ptx --dflash-n-draft must additionally be ≤ 3.
• Tool calling — --tool-parser atem converts Muse Glimmer's ATEM output into OpenAI tool_calls. Choices are atem or none; the default is none, so pass it explicitly if you want tools.

Smoke test

curl http://127.0.0.1:8000/v1/chat/completions \
  -H 'Content-Type: application/json' \
  -d '{"model":"muse-glimmer-30B","messages":[{"role":"user","content":"What is the capital of France?"}],"max_tokens":32,"temperature":0}'

Endpoints: GET /health, GET /v1/models, POST /v1/chat/completions (streaming supported), plus DELETE /v1/sessions/{id} and POST /v1/sessions/{id}/reset for named sessions.

---

Gotchas

Filenames. <dirname>.pte / <dirname>.ptd, not model.pte / aoti_cuda_blob.ptd. Both --model-path and --data-path differ between the download path and the local-export path. Any guide written against a local export needs both paths translated.

A CUDA .pte alone is inert. 16–35 MB of graph with no weights. Bring the .ptd.

One worker only. --num-runners must be 1; more would duplicate the weights in memory. The server errors out otherwise.

Stop tokens. The turn ends on <|end_of_text|> or <|eot|>. Never stop on <|eom|> — it ends a single message and continues the turn. On a tool call the model emits a private reasoning message closed by <|eom|> first; treating it as terminal means the tool call is never generated. The bundled runtime handles this; it matters if you write your own client against the raw runner.

Unsupported OpenAI parameters return a structured 400 rather than being silently ignored, including reasoning_effort and tool_choice="required" (only none / auto / unset are accepted). The server honors temperature, top_p, top_k, seed, max_tokens / max_completion_tokens, stop, and tools. To control reasoning depth, use the system prompt (Reasoning strength: <value>), not reasoning_effort.

Sampling defaults. temperature 1.0, top_p 0.95, top_k 64.

---

Runtime limitations

These are limits of the ExecuTorch runtime and this server, not of the model — for the model's own limits see Considerations and Limitations below.

• No video input. Text and images only, one image per request, JPEG or PNG. There is no video or frame-sequence path.
• No continuous batching. One request executes at a time. --num-runners must be 1, every exported method is batch-1, and execution is serialized behind a single lock. Named sessions give you isolation between conversations, not concurrent throughput.
• No cross-session prefix sharing and no checkpointing. Each session owns its own KV cache, nothing is reused across sessions, and session state is never persisted — POST /v1/sessions/{id}/reset zeroes it rather than saving it. --warm-resume is a within-session optimization and does not cross a session boundary.

The PyTorch team lists all three as work in progress in the announcement.

---

Muse Glimmer-30B Model Overview

Building effective agents requires key capabilities working together to achieve the user's goals. Muse Glimmer is trained and evaluated on these capabilities:

• End-to-end Agentic Task Completion. Muse Glimmer achieves strong success rates on full-task benchmarks including DeepSearch QA, MCP-Atlas, 𝛕3-Bench and SWE-Bench, which measure its ability to work within scaffolds, write and debug code, and resolve multi-turn requests from start to finish.
• Reliable Tool Use. The model handles a wide range of function calls, invoking tools with precise schemas throughout extended workflows.
• Multi-Step Reasoning. Muse Glimmer chains reasoning over long horizons, sustaining coherent plans across complex, extended workflows.
• Failure Recovery. When a tool call fails or returns an unexpected result, the model diagnoses the error and retries rather than halt.
• Multimodal Input and Reasoning. Through a dedicated perception encoder, the model accepts interleaved text and images. This enables agents to interpret screenshots, charts, and documents alongside conversation.
• Scaffold Compatibility. Muse Glimmer works across OpenClaw, Hermes Agent, and other agentic orchestration patterns.
• Controllable Effort. The model supports different reasoning strengths to select the right balance between quality and speed.
• Multilingual. Muse Glimmer is trained on data from more than 100 languages.

Optimized for Local Deployments

Muse Glimmer was optimized for local deployment, and designed to run at practical speeds on consumer hardware without sacrificing quality.

Fitting the Model on Your Device. We use quantization techniques to compress the model's weights to approximately 4-bit precision, shrinking the language model to under 20 GB. This leaves enough headroom for the model's KV cache, the perception encoder for image understanding, and the speculative decoding drafter to run simultaneously within a 24 GB or 32 GB envelope. Critically, we validated that this compression introduces minimum to no degradation on agentic tasks.

* Degradation measured using an average on accuracy metrics across 15 common benchmarks

Faster Generation Through Speculative Decoding. Muse Glimmer ships with a lightweight "drafter" model based on DFlash, a small companion network that proposes entire blocks of tokens at once. The DFlash block-diffusion model predicts entire blocks of 16 tokens in a single forward pass. The main model then verifies these proposals in parallel, accepting correct tokens and correcting wrong ones. This technique lets Muse Glimmer generate text significantly faster than standard token-by-token generation while producing identical output quality. We provide quantized drafter versions to incur a smaller memory overhead in the release. This is what the dflash variants in this repo contain.

We measure the speed of our K-Quant-17GB model alongside the quantized DFlash drafter on MacBook M4-Max, M5-Max and on an Nvidia RTX-5090. The model is fast enough for fluid conversation and real-time agent interaction, all running entirely on your device.

* Average across a diverse prompt set. Measurements done with batch size 1 and greedy decoding. M4/M5 measurements were done using ExecuTorch, and RTX using llama.cpp.

Benchmarks

We evaluated Muse Glimmer across a broad range of benchmarks to assess the diverse capabilities required for effective autonomous agent behavior. Compared with Gemma4-31B and Qwen3.6-27B, Muse Glimmer performs strongly for its size class on several widely used LLM benchmarks.

For more detail about our evaluations, see our report.

Best Practices

To achieve best performance, we recommend the following settings:

Sampling Parameters: Use the following configuration:

• temperature = 1.0
• top_p = 0.95
• top_k = 64

Reasoning Strength: Reasoning strength controls how much the model thinks before responding to the prompt. Reasoning strength can be defined as part of the system prompt as Reasoning strength: <value>. Muse Glimmer supports the following levels: low / medium / high / xhigh. Use high or xhigh for complex problem solving, coding, and agentic tasks. The OpenAI reasoning_effort request parameter is not supported by the ExecuTorch server and is rejected with a 400 — use the system prompt.

Trust and Safety

As we would for other large language models, we strongly recommend that Muse Glimmer be deployed not as an endpoint in itself but as part of an overall AI system with additional guardrails as required or appropriate for the use cases and context of its deployment. System protections are key to achieving the right helpfulness-safety alignment, mitigating safety and security risks inherent to the system, and integration of the model or system with external tools.

Evaluations
We evaluated Muse Glimmer for common use cases as well as specific capabilities. Common use cases evaluations measure safety risks of systems for most commonly built applications including chat bot and visual, QA. We built dedicated, adversarial evaluation datasets and evaluated systems composed of Muse Glimmer models and those safeguards to filter input prompt and output response. It is important to evaluate applications in context, and we recommend building dedicated evaluation datasets for your use case.

Capability evaluations measure vulnerabilities of models inherent to specific capabilities, for which were crafted dedicated benchmarks. We also used industry standard safety and capability benchmarks where appropriate.

Muse Glimmer was primarily evaluated across four risk axes:

1. Content safety — Standard alignment for refusal of harmful requests and calibrated responses to borderline prompts.
2. Agentic risk — Policies for irreversible-action confirmation, data minimization, scaffold boundary respect, and indirect prompt-injection resistance.
3. Privacy (Appropriate Information Flows) — Respect for contextual integrity of information when interacting with third parties on an individual's behalf, inspired by CI theory.
4. Preparedness — Chemical & biological, cyber, and loss-of-control risks.

Preparedness

Muse Glimmer does not fall under the definition of "Frontier AI" in Meta's Advanced AI Scaling Framework (AAISF), since it is generally less capable than Muse Spark. However, as a matter of prudence, our Preparedness Team assessed Muse Glimmer's risk profile and determined that it would receive the following designations:

• Chem/Bio: Moderate or lower risk;
• Cyber: Moderate or lower risk (inferred);
• Loss of Control: Moderate or lower risk (inferred).

Cyber and Loss of Control risk levels are inferred to be Moderate or lower since Muse Glimmer is broadly weaker than Muse Spark 1.0, which received the same risk designation in these domains.

In the chem/bio domain, we evaluated Muse Glimmer on a range of benchmarks for scientific knowledge and wet-lab debugging (most performant in Muse Glimmer's size class are bolded; second most performant is underlined — Kimi K3 is also included for context):

We find that Muse Glimmer's abilities are approximately in line with other models in its size class, while showing strictly lower capabilities than larger open-weight models, suggesting that it is unlikely to materially enable new threats upon release. We also evaluated it on our suite that focuses on the unique set of bottlenecks that would otherwise deter or limit the success of real-world threat actors; here, our evaluation rated its risk rating at moderate or lower as well. See the Muse Spark Safety & Preparedness Report for a detailed description of the above evaluations and our methodology.

Train-Time Mitigations

1. Safety SFT: Curated examples demonstrating correct safety behavior, including agentic safety scenarios covering tool-use boundaries, prompt injection resistance, and permission handling.
2. Safety RL: Reinforcement learning with safety-specific reward signals that penalize policy violations while rewarding helpful responses to legitimate requests.
3. Appropriate information flows: Principles of data sensitivity recognition, minimization, and local-first execution embedded directly into model weights through dedicated synthetic training data.

Intended Use

Intended Use Cases: Muse Glimmer is intended for commercial and research use. The model is optimized for autonomous agentic tasks including:

• Local AI agents: Multi-step planning, sequential tool invocation, failure recovery, and long-horizon task execution running entirely on consumer devices.
• Coding agents: Writing, debugging, and resolving real-world software engineering tasks (e.g., SWE-Bench style workflows).
• Tool use and function calling: Reliable schema-based tool invocation across extended, multi-turn workflows.
• Multimodal reasoning: Interpreting screenshots, charts, documents, and images alongside conversation for agentic and information-rich environments.
• Synthetic data generation: Generating high-quality training data for downstream model development.
• LLM-as-a-judge evaluation: Serving as an evaluator for other models' outputs.

Out-of-scope: Use in any manner that violates applicable laws or regulations (including trade compliance laws). Use in any other way that is prohibited by the Apache 2.0 License terms. Audio input/output is not supported.

Considerations and Limitations

Muse Glimmer is a technology that carries known and unknown risks. Testing conducted to date has not, and could not, cover all scenarios.

Limitations:

• The model may produce inaccurate, biased, or objectionable responses to user prompts.
• While optimized for agentic tasks, the model may still make errors in multi-step reasoning, particularly in novel scenarios not well represented in training data.
• The model is not explicitly optimized for video; video input is processed as individual frames.
• The model has not been evaluated on all languages contained in the pre-training data. Performance may degrade on languages outside the strongly supported set.
• Quantized inference may show minor quality differences in edge cases compared to full-precision.
• The model is not intended to be downloaded by or used by individuals under the age of 18. Where deployed within systems that may be used by individuals under the age of 18, deployers are responsible for ensuring that any risks associated with such use by individuals under the age of 18 has been fully assessed and appropriately mitigated, and complies with all applicable laws.

Responsible Use: Developers should perform their own safety testing and tuning tailored to their specific applications and proposed languages. Our Usage Policy can be found here [link]. We recommend implementing additional guardrails (such as human-in-the-loop confirmation for irreversible actions) when deploying the model in agentic contexts where it can take real-world actions.

Released Artifacts

All artifacts are released under Apache 2.0:

Related repos: meta-models/Muse-Glimmer-30B (BF16 weights and tokenizer metadata) and meta-models/Muse-Glimmer-30B-GGUF (GGUF checkpoints, and the input format for exporting your own PTEs).

Where to send questions or comments about the model: Please provide any feedback, comments or bug reports on the model through the Hugging Face page at https://huggingface.co/meta-models/. For more technical information about generation parameters and recipes for how to use Muse Glimmer in applications, please see the developer documentation.

Further reading

• Fast on-device agentic AI with ExecuTorch — the PyTorch team's announcement post, including their performance measurements. Note that its quickstart is written against a local export and uses model.pte / aoti_cuda_blob.ptd; use the filenames on this page for the pre-exported artifacts.
• Muse Glimmer in ExecuTorch — export from GGUF, custom quantization recipes, exported method contracts, and runtime internals.