Meta Printed the 13 Rows Its Model Lost

An 82 percent guess rate, a 24GB fit, and a flagship going open.

Meta Superintelligence Labs released Muse Glimmer on August 10, 2026, under Apache 2.0. It compares the model against two peers across 26 published metrics. Muse Glimmer leads 13 of them.

The other 13 are in the same table. Meta printed them, named which model led each row, and published a methodology report explaining where its own harness was generous to competitors. That is a rarer document than the model.

This report walks the ledger item by item: what shipped, which rows split which way, what an independent lab measured separately, and which parts of the trust story remain yours to build. Every claim below carries a stamp for whether the vendor, an outsider, or nobody has verified it.

A 29.6B-parameter dense transformer with a built-in vision encoder, distilled from the closed Muse Spark flagship and tuned for agent loops that fit one consumer GPU. It is Meta's first open-weights release since Llama 4, sixteen months ago, and its first under Apache 2.0 rather than a Llama community license.

The architecture is built around a memory budget rather than a leaderboard. Fifty-two decoder layers, model width 6656, a 16:1 grouped-query ratio, and a repeating pattern of three sliding-window layers (window 2048) per global layer. That pattern is why the KV cache settles near 1.8GB at 128K context instead of eating the headroom the weights left behind. Context is listed at 131,072 tokens with extension, the tokenizer carries 202,048 entries, and the knowledge cutoff is January 2026.

Released artifacts, all Apache 2.0: BF16 weights at 59.6GB, two 4-bit quantized variants, the frozen 1.8B perception encoder, and a DFlash speculative-decoding drafter. A separate usage policy sits beside the license in the same repository.

Read that footnote against the sentence above it. The body text attaches the claim to agentic tasks, saying compression costs minimum to no quality there. The footnote says the figure is an average of accuracy metrics across 15 common benchmarks. Those are different populations. A 1.0 percent average accuracy drop is a per-answer measurement, and agent runs in these harnesses stretch to 200 turns and beyond, where a small per-step regression compounds into a failed trajectory. That reading is ours, not Meta's claim, and it is the first thing worth reproducing on your own workload.

The speed story is cleaner. DFlash is a five-layer block-diffusion network that proposes 16 tokens per forward pass, which the main model verifies in parallel. On the K-Quant-17GB build, Meta measures 74.9 to 233.4 tokens per second on an Nvidia 5090 (3.1x), 26.6 to 50.2 on an M5 Max (1.8x), and 23.7 to 37.8 on an M4 Max (1.5x), at batch size 1 with greedy decoding, using llama.cpp on the Nvidia card and ExecuTorch on the Macs.

Meta compares Muse Glimmer against Gemma4-31B and Qwen3.6-27B across 26 published metrics. By our count of that table, Muse Glimmer leads 13, Qwen3.6-27B leads 9, and Gemma4-31B leads 4. The launch post calls this performing strongly for its size class, which is the accurate phrasing, and it is worth noticing Meta did not reach for a stronger one.

Where the wins land matters more than the count. Muse Glimmer takes the orchestration rows. Qwen takes the rows where an agent touches the operating system or the screen, plus three of four multimodal rows. Gemma takes both lower-is-better safety rates. That is a workload map, not a ranking.

Artificial Analysis had pre-release access and published its own run on launch day. Muse Glimmer scores 35 on the Intelligence Index, against 38 for Qwen3.6 27B and 36 for Kimi K2.5, with Gemma 4 31B behind it on every measure they report. Median for comparable models is 9, so the placement is strong for the size class.

The gaps concentrate in the category Meta named as the product. On GDPval-AA v2, which scores realistic knowledge work inside an agentic loop, Muse Glimmer records 953 Elo against 1141 for Qwen3.6 27B. On Terminal-Bench v2.1, 52 percent against 61. And on knowledge calibration, an AA-Omniscience Index of minus 33, driven by an 82 percent hallucination rate against 49 percent for Qwen.

The exception deserves as much volume as the gaps. On Tau3-Banking, a tool-use evaluation over unstructured policy documents, Artificial Analysis puts Muse Glimmer at 24 percent, ahead of Gemini 3.5 Flash-Lite at 18 and Qwen3.6 27B at 17. Meta's own table agrees. Structured tool calling against a schema is the thing this model does better than its class, and it is the thing local agents do most.

The evaluation methodology report is the most useful document in the release, and it is candid in a direction that costs Meta points. Four disclosures worth carrying forward.

Meta also states plainly that its harness was not tuned for third-party models, so competitor rows may understate their best performance. Taken together, the report tells you the table is a snapshot of three models inside one scaffold, which is the correct way to read every agentic leaderboard published this year.

A 28.4 percent attack success rate is not scandalous in isolation. Every model in this class is weak here: Gemma sits at 25.6, Qwen at 40.3, and the attacker in that benchmark is a frontier model refining its injection up to six times across 949 scenarios. What changes is the scaffolding around the model at inference time.

When the same agent runs behind a hosted API, the provider typically contributes an injection classifier, request logging, rate limits and an abuse team. Run the weights on your own machine and every one of those disappears at the moment the model gains a shell, a browser, a mail client and your file system. The security boundary becomes the harness, the loopback runtime and the tool allowlist you wrote yourself. Muse Glimmer's utility score of 94.2 sharpens this rather than softening it: it keeps completing the user's task while the injected task also succeeds.

Meta says as much in the model card. It recommends deploying the model as part of a system with additional guardrails, and specifically recommends human confirmation for irreversible actions in agentic contexts. That guidance is correct, and it is also the entire distance between the demo and a deployment.

The weights landed beside a 6,500-word letter from Mark Zuckerberg, The Future is for Everyone. The headline commitment is not about Glimmer at all. Zuckerberg wrote that Meta will "soon release the weights for Muse Spark 1.2", its current flagship, reversing the closed-weights posture that governed the first three Muse Spark releases.

Treat that as stated, not shipped. There is no repository, and the date, parameter count and license are unconfirmed. It matters anyway, because Muse Spark 1.2 scores 54 on the Artificial Analysis Intelligence Index, above every US open-weight model you can download today. If it lands, the model you are evaluating this week is superseded within your own datacenter, by the same vendor.

Two further load-bearing lines for engineers. On governance, Meta is implementing a structure giving its independent board the power to approve safety criteria for model releases and review whether each release meets them. On alignment, Zuckerberg rejects a centrally defined value set in favor of agents aligned to each user's goals, and argues trust is the adoption gate: people will not hand sensitive tasks to an agent they believe serves the vendor. That argument is also the commercial case for the model in section 01.

Four points that cut against the critical read, stated at full strength.

Benchmark the quantization you will actually ship. The published table is a high-reasoning run with quantization losses reported separately, so treat those numbers as the unquantized ceiling. Your build is a 4-bit variant with a drafter attached. Measure your own long-horizon task completion at K-Quant-17GB before trusting the 1.0 percent.

Treat the harness as the security boundary. The model contributes 28.4 and 26.4. Everything better comes from your tool allowlist, workspace scoping, loopback binding and a confirmation step before any irreversible action, which is Meta's own recommendation.

Assume overconfidence and design for it. An 82 percent guess-rather-than-abstain profile means retrieval grounding and a verification step matter more here than with a hosted frontier model. Reserve it for tasks with a checkable output.

Match the workload to the strength. Schema-driven tool calling and multi-turn orchestration are where it leads. Desktop control, terminal work and document parsing are where Qwen3.6-27B leads in Meta's own numbers.

Read both documents, and watch for the flagship. Apache 2.0 governs the weights, and a separate usage policy ships in the same repository. Training data and the full pipeline are not released, so open weights is the precise term for what you have. If Muse Spark 1.2's weights arrive as stated, re-run your evaluation against it before committing a year of tooling to the 30B.