AI Brief, 16 October 2026: the qualifier doing the work in Meta's open-weights claim

Meta released the weights for Muse Spark Open on 15 October: a natively multimodal mixture-of-experts model with 816 billion total parameters and 42 billion active, a 1,048,576-token context, and a bespoke licence Meta calls the Muse Model License. The announcement headline calls it "the most capable US open-weight model to date", and that qualifier is carrying the sentence. This brief asked on 3 September whether Meta would ship the Muse Spark weights at all, after a post committed to "the Muse Spark open weights release" with no date. Six weeks later there is an artifact.

The evidence for the capability claim is harder to use than it should be. Meta's comparison table is published as a single rasterised image with no numbers in text, the second Meta launch in four days to ship its benchmarks as a picture. Read off that image, the model leads on six of eleven rows. On the two rows testing what the post says it was built for, it does not: on Terminal-Bench 2.1 it scores 76.4 against 85.0 for Moonshot's Kimi K3, 83.9 for Z.ai's GLM 5.3 and 82.0 for Alibaba's Qwen3.8-2.4T-A95B, and on DeepSWE 1.1 it scores 60 against 68 and 67 for Kimi K3 and GLM 5.3, beating Qwen3.8 there at 57. Every model ahead of it on either row has public weights on Hugging Face, and every one is Chinese. Remove the word "US" and the claim does not survive its own scorecard.

An independent reading points the same way. Artificial Analysis scored the model on 15 October and places it at 38.87 on its Intelligence Index, against 48.1 for Muse Spark 1.3 and 39.6 for Muse Spark 1.2 — and all three cost the same, $1.25 per million input tokens and $4.25 per million output. Across individual evaluations it tracks the August model closely and the September one poorly. Meta's own announcement assets are named muse-spark-1-2-open-weights, which agrees with what the scores show: the weights released are the lineage of Muse Spark 1.2, not 1.3.

The licence is the other half of the release, and it is not an open-source licence. The grant is shaped like MIT, but any company turning over more than twenty million dollars in gross revenue must negotiate a separate written agreement with Meta before serving the model to anyone else.

• Muse Spark Open: 816B total, 42B active, 1M-token context, multimodal, released 15 October; weights gated behind a form requiring full legal name, date of birth and affiliation.
• Meta's benchmark comparison exists only as an image. Read off it, the model sits behind Kimi K3, GLM 5.3 and Qwen3.8 on Terminal-Bench 2.1, and behind the first two of those on DeepSWE 1.1.
• Artificial Analysis independently scores it 38.87, below Muse Spark 1.3 at 48.1 and level with Muse Spark 1.2 at 39.6, at identical pricing.
• On Terminal-Bench 4.0 it scores 8.1%, where Muse Spark 1.3 scores 33.3%. Meta's scorecard reports version 2.1 of that benchmark, not 4.0.
• The licence's commercial trigger is $20M of the licensee's total gross revenue, not revenue from serving the model.
• Microsoft AI disclosed that MAI-Base-2, 2.8 trillion parameters with 80B active, was pre-trained on 16,000 GB200 GPUs over 38 days at 21% average utilisation.

The scorecard, read off the picture #

Meta's "Key Benchmarks" section contains no table and no figures in text: it is one PNG, in light and dark variants. The same pattern was noted here on 13 October for the Muse Wonder launch, where the missing table meant a state-of-the-art claim could not be checked. The figures below are read from that image.

Two caveats are visible in the image itself. The comparison runs models at different reasoning efforts: Muse Spark Open and Qwen3.8 at xhigh, GLM 5.3 and Kimi K3 at max. And the DeepSWE 1.1 row is annotated "Best available/ Self reported harness", so the competitors' figures there are lifted from their own published runs rather than measured on a common harness. The Terminal-Bench 2.1 row carries no such annotation.

The more consequential question is which version of Terminal-Bench appears. Meta reports 2.1. Artificial Analysis's index runs Terminal-Bench 4.0, where Muse Spark Open scores 8.1%. For scale, the same board gives Muse Spark 1.3 33.3%, DeepSeek V4.1 Flash 26.8%, Kimi K3 12.6% and Muse Spark 1.2 7.1%. A figure that low usually indicates the harness rather than the weights — an agent that cannot reliably drive a tool loop scores near zero however good its code is — so read it as an open question about scaffolding, not a verdict on the model. But it is the number a reader evaluating this for agentic work needs, and not the one the launch gives.

One honest complication: on OpenRouter the open model accepts audio input, which neither Muse Spark 1.2 nor 1.3 does. So the weights are not simply 1.2 re-published, and Meta has not said what they are.

What you actually get when you run it #

The licence is worth reading in full: its shape is unusual and its hard threshold far tighter than Meta's previous open releases. The grant is close to MIT — use, copy, modify, fine-tune, create derivative works, distill, merge, publish, distribute and sell — and Meta "claims no ownership of outputs".

Then four conditions. Copies must carry the notice and the Usage Policy, which prohibits military, nuclear, espionage and ITAR-controlled applications and operation of critical infrastructure. Any product built on the model must display "Built with Muse Spark Open", and "Derivatives" is defined broadly enough to cover any model distilled from its outputs. The condition that decides who can use this commercially is the third:

... where the third party receives outputs of such models ("Model as a Service"),
and the aggregate gross revenue of the Licensee and its affiliates exceeds
twenty (20) million US dollars ... during any consecutive twelve (12) month
period, the Licensee must notify Meta and enter into a separate written
agreement with Meta before offering the Software ... as part of that Model as a
Service business.

Read the subject of that threshold carefully. The twenty million dollars is the licensee's and its affiliates' aggregate gross revenue, not revenue earned from serving the model. A company with $50M of unrelated revenue and $100k of inference revenue is captured. The carve-outs are real but narrow: internal use, research, and powering features of a product that is not itself a model-serving business. Llama's equivalent trigger was 700 million monthly active users, which perhaps a dozen companies on earth crossed. Twenty million dollars of revenue is a mid-sized business.

Access is gated as well: downloading requires full legal name, date of birth, country, affiliation and job title, with IP-based geolocation recorded, under manual approval. The model card, configuration and weights all sit behind that form. A day after release the repository showed 291 likes against 402 downloads, the profile of a release being watched rather than run.

Where the openness pays off immediately is in what third-party serving exposes. On OpenRouter, Muse Spark Open is the only Meta model listed that is not marked moderated, and the first to expose logprobs, top_logprobs, seed, logit_bias, min_p and the frequency and presence penalties. Every other Meta listing offers temperature, top-p, top-k and reasoning effort and nothing else. That difference is not a licence term; it is what happens when the people serving a model are not the people who trained it, and for constrained decoding, calibration or reproducible sampling it is the most useful thing here.

One number deserves arithmetic rather than agreement. Meta shipped auto-tunable Helion GPU kernels into SGLang and reports "1.03-3.95× kernel-level speedups" alongside "a 3% end-to-end speedup on AMD MI350". Let $f$ be the fraction of end-to-end serving time spent in the replaced kernels and $s$ the speedup on that fraction. Amdahl's law gives the overall speedup as

$$S=\frac{1}{(1-f)+f/s}$$

Take the best case Meta reports, $s=3.95$ , with the claimed outcome $S=1.03$ . Solving for $f$ gives $f\approx 0.039$ : granting the top of the range to every replaced kernel, those kernels were about four percent of end-to-end time, and since most of the range sits nearer 1.03x, four percent is a floor.

def amdahl_fraction(overall_speedup, kernel_speedup):
    """Share of runtime the optimised kernels must occupy to explain an observed win."""
    return (1 - 1 / overall_speedup) / (1 - 1 / kernel_speedup)

amdahl_fraction(1.03, 3.95)   # 0.039 -> kernels are ~4% of runtime, at best
amdahl_fraction(1.03, 1.50)   # 0.087 -> at a realised 1.5x, ~9%

Both figures are Meta's own and unreproduced. "Up to 3.95x faster" and "3% faster" describe the same change, and only the second is what a reader deploying this model will observe.

A researcher retracts his own evidence for superposition #

The most-discussed AI item of the day came from neither a frontier lab nor a paper. Nick Cammarata, now at Goodfire, previously a co-author of Distill's Zoom In and lead of OpenAI's first large-scale automated-interpretability project, published an essay arguing that neural networks are already interpretable and that the canonical evidence for superposition — including a circuit he produced himself — was misread. It took 564 points and 216 comments on Hacker News on 15 October, the day it went up.

The technical claim is specific. Superposition holds that networks pack more features than they have neurons, storing them in non-orthogonal directions, so neurons are polysemantic by necessity and must be decomposed by a learned dictionary such as a sparse autoencoder before anything can be read. Cammarata's counter-claim is that the polysemanticity seen in early model organisms came from ordinary, fixable causes — chiefly an architectural failure in deep layers — not from a second learning principle. Retraining InceptionV1 with modern deep-layer fixes, with no change to model size or training data, he reports 99-100% of last-layer neurons easily interpretable, against almost none originally. A preliminary look at Qwen3-32B puts roughly 90% of neuron poles on one main job and 2-4% not yet understood. The baselines are stark: Anthropic's 2023 Towards Monosemanticity scored its median neuron 0 out of 14 on its own rubric, and OpenAI's 2023 automated interpretability work found about 1.7% of neurons interpretable by its most generous measure.

The epistemic status needs saying plainly, because the essay says it itself: this is labelled "an informal research essay and working hypothesis". The promised technical companion paper is not out. The language-model percentages are preliminary internal estimates rather than audits, and the essay says full audits at trillion-parameter scale have not begun. The most credentialed dissent is already on the record: Neel Nanda, who leads interpretability at Google DeepMind, called the work interesting but declined to treat it as a significant update to the big picture of interpretability. A commenter made the sharper methodological point that axis-aligned semantic directions are not what the superposition hypothesis predicts in the first place, so the result may be answering a weaker claim than the one being contested.

It is reportable regardless: a researcher with standing in this literature is withdrawing confidence in evidence he generated himself. That is rare enough to be worth attention even before anyone reproduces it, which nobody yet has.

Also notable #

• Microsoft AI disclosed how it trained MAI-Base-2 on 15 October, and the disclosure is more interesting than the model: 2.8 trillion parameters with 80B active, 50T training tokens, 16,000 GB200 GPUs over 38 days, 21% average MFU and 89% goodput, and fitted scaling laws published in the open. That specificity is rare. The widely repeated "4-12x more pre-training FLOPs-efficient than DeepSeek-V4-Pro-Base" is not a model-to-model measurement: it is Microsoft's estimate of what its recipe would need to match DeepSeek's bits-per-byte on four of its own held-out evaluations, read off scaling-law fits that include MAI-Base-2 itself, and the 4.2-12.3x range is the spread across those four evals, not any interval. Microsoft never states the DeepSeek compute figure the ratio divides by, and concedes the reference "is still trained on significantly fewer FLOPs". The figure carrying the claim is, again, a rasterised image. The model is in post-training; no weights. At 37 points on Hacker News, essentially nobody is covering it.
• Liquid AI's LFM2.5-24B-A2B, published 13 October and unreported here at the time, is 24B total with 2.45B active, its 40 layers running a fixed 3:1 schedule of 30 short-convolution to 10 full-attention blocks, with a 131,072-token context after 36T training tokens. Liquid reports 184 tokens per second on an Apple M5 Max at Q4_K_M. Weights are downloadable under a custom lfm1.0 licence, not an OSI one; every benchmark figure is vendor-run, and the agentic wins come from a recipe doing reinforcement learning inside the harnesses being scored.
• NVIDIA's Nemotron 3.5 Super VL, a 120B open multimodal model, reached OpenRouter at 15:04 UTC on 15 October at $0.15 and $0.50 per million tokens with a free tier. Weights landed 12 October; this is the first hosted availability.
• Krea 3 launched as API-only, from a studio that shipped open weights twice in sixteen months. Its speed comparisons are self-run against pre-generated outputs.
• Anthropic published an update on independent evaluators, describing employee-level access for METR researchers. Governance, not capability.
• Training-free on-policy distillation (arXiv 2610.22236, 15 October) argues that merging separately fine-tuned specialists needs no gradient steps: when specialists share a base, each linear layer has an exact least-squares solution needing only the second moments of that layer's inputs, from one forward pass over a few hundred rollouts. Shown on Olmo 3 7B, self-reported.

What to watch #

• Whether anyone states what Muse Spark Open actually is. The launch names no base model and no training recipe. The asset naming says 1.2, the independent index agrees with it to within 0.7 points, and the audio modality says otherwise. A model card with a lineage would settle a question currently resting on filenames.
• Whether a second party reproduces the Terminal-Bench 4.0 figure of 8.1%. The weights are downloadable and the harness public, so anyone through the gate can check it. If it holds, the gap to Muse Spark 1.3's 33.3% is the most important number about this release; if it is a scaffolding artifact, it should move a long way.
• Whether Cammarata's companion paper arrives with the language-model audits in it. The InceptionV1 result has a browsable artifact behind it; the Qwen3-32B and trillion-scale figures, the ones that would matter to practitioners, do not.
• Whether any decision-model vendor publishes a fail-open rate. Asked here on 10, 11, 12, 13, 14 and 15 October, now seven days running, with no vendor answer yet.