Muse Spark gated at $20M, Antigravity skips MIC data, FLUX.3 Klein adds 28GB

TL;DR

• Meta’s Muse Spark gates commercial use above $20M revenue in its new 816B MoE release.
• Muse Spark’s NVFP4 path breaks day one with SGLang dropping weights and vLLM crashing on SM120.
• Google Antigravity flags a bacteriocin candidate with no MIC data or false-positive rate published.
• FLUX.3 Klein’s mandatory prompt upsampler adds ~28GB VRAM on top of the 23.8 GiB base.
• Klein scores 24/32 on instruction checks versus GPT Image 2.5’s 31.

Three releases land in the tech beat today, and each one asks the reader to look past the operational detail it buried under the launch copy. Meta’s Muse Spark is an 816B MoE with a 1M-token multimodal context — and a $20M revenue gate in a new license, and an NVFP4 deployment path that doesn’t actually run on SGLang or vLLM yet. Google Antigravity flagged a novel bacteriocin candidate — without the MIC data, false-positive rate, or head-to-head numbers that would make the claim checkable. Black Forest Labs’ FLUX.3 Klein pitches lean open-weights image generation — until you notice the mandatory prompt upsampler adds roughly 28GB VRAM on top of the base model.

The three stories don’t share a vendor, a modality, or a benchmark. They share a pattern: the launch pitch and the deployable reality travel in separate documents, and only one of them makes it into the blog post. The round-ups — Nvidia’s Indic pretraining pipeline, IBM’s Granite Switch 4.1 unified endpoint, and Simon Willison’s audio-chirp Pong — sit alongside as the day’s cleaner shipping stories.

Meta’s 816B Muse Spark Open ships gated above $20M revenue

Source: huggingface-blog · published 2026-10-15

TL;DR

• 816B total / 42B active MoE from Meta, 1M-token context shared across text, image, audio and video.
• Artificial Analysis scores it 39, behind Kimi K3 (44) and MiMo-V2.6-Pro (46) on the Intelligence Index.
• $20M revenue gate in the new Muse Model License forces separate commercial terms with Meta.
• NVFP4 deployment is broken day one: SGLang silently drops layer weights, vLLM crashes on SM120.

What shipped

Muse Spark Open is Meta’s first frontier-scale MoE release: 816B total parameters with 42B active per token, routed 8-of-256 experts per layer across a 62-layer, 8192-dim backbone with grouped-query attention (128 Q heads, 16 KV). A 1.8B perception encoder handles images and video (14×14 patches, 2 FPS up to 96 frames); a convolutional audio pyramid emits 12.5 tokens/sec from 24 kHz mono. The context window is 1,048,576 tokens, shared across modalities — about 266 GB of KV cache at BF16 for a full fill.

The agentic framing is the pitch. The “Harmony” chat template splits reasoning into self, commentary, and user channels, with a four-level reasoning-effort knob (low → xhigh). Weights ship in BF16 (~1.5 TB) and NVFP4 (4-bit, Blackwell-tuned); Meta claims the 816B model fits on a single B200 node. TRL fine-tuning with LoRA on the attention projections — experts frozen — moved mean token accuracy from 72% to 84% over 500 steps on 64 H100s.

Benchmarks don’t match the marketing

Meta’s “most capable US open-weight model” framing is the single most-contested claim around the launch. Independent evaluation from Artificial Analysis places Muse Spark Open at 39 on the Intelligence Index — behind Moonshot’s Kimi K3 (44) and Xiaomi’s MiMo-V2.6-Pro (46) — and DeepSeek V4.1 Flash beats it on Terminal-Bench 4.0 by 27% to 8% ¹. The “US” qualifier is doing a lot of work.

The published Terminal-Bench numbers themselves are under fire. Hacker News commenters allege Meta’s harness used 6 CPU cores and 8 GB of RAM — well above the task’s official resource limits — and called the resulting scores “cheating” ². François Chollet publicly argued the model is over-fit to public benchmarks, and Meta’s own Alexandr Wang conceded it fails baseline ARC-AGI-2 requirements ³. A week-long hands-on test reported Muse Spark fumbling tasks as pedestrian as Mac swipe gestures and Three.js physics setups that DeepSeek v4.1 completes in one pass ⁴.

”Open weight” with a revenue gate

The license change is the structural story. The bespoke Muse Model License replaces Llama’s community terms with a $20M annual-revenue threshold: cross it, and you must notify Meta and sign a separate commercial agreement before deploying the model ⁵. Critics argue this fails OSI Freedom 0 outright and functions as regulatory arbitrage against the EU AI Act’s documentation requirements for proprietary systems ⁵. The optional “Contributor tier” — discounted inference in exchange for data-sharing rights — has been described by early reviewers as “creepy” ⁴.

A $20M annual-revenue threshold requires separate commercial negotiation with Meta, which commenters argue violates OSI Freedom 0. ⁵

NVFP4 deployment isn’t ready

The “single B200 node” talking point assumes an inference stack that doesn’t yet exist cleanly. SGLang silently drops weights for specific layers when loading llm-compressor NVFP4 checkpoints, and vLLM’s --kv-cache-dtype nvfp4 flag triggers EngineDeadError on SM120 because the required TensorRT-LLM kernels are missing ⁶. llama.cpp offloading of NVFP4 tensors to CPU risks functional incorrectness, and speculative-decoding paths hit weight-shape mismatches under the new quant layout ⁶. The headline deployment story will probably hold in a quarter; today, it ships with footguns.

The takeaway: Muse Spark Open is a real engineering artifact at a scale the open ecosystem hasn’t seen from Meta before. It is not the frontier leader, it is not open-source, and the NVFP4 path isn’t production-ready. Read Meta’s launch post with those three caveats in hand.

---

Google Antigravity flags a bacteriocin lead, not a drug

Source: deepmind-blog · published 2026-10-15

TL;DR

• Google Antigravity flagged a 97-aa peptide (UniProt A0A0U0D591) in S. pseudopneumoniae as a novel Lactococcin 972-family bacteriocin candidate.
• ProtNLM, the model doing the work, has annotated 49M+ TrEMBL entries and scored 79% correct in EMBL-EBI’s blind curator review.
• DeepMind published no MIC data, no false-positive rate, and no head-to-head against incumbent AMP predictors like AMPlify or OmegAMP.
• Independent reviewers of Antigravity 2.0 report corrupted installs and a UI that hides warnings once the agent deems code safe.

What the demo actually shows

DeepMind’s new blog pitches Google Antigravity as an “agentic science workbench”: a Gemini 3.8 Flash agent that queries UniProt, runs NCBI-BLAST, pulls AlphaFold structures, and reasons over genomic neighborhoods in one session. The headline result is a 97-amino-acid peptide (UniProt ID A0A0U0D591) from Streptococcus pseudopneumoniae, flagged as a novel member of the Lactococcin 972 family. The agent noted the candidate sits 6 bp upstream of an ABC transporter operon — strong contextual evidence for a functional bacteriocin — and that its AlphaFold fold matches experimentally validated family members.

That is a biologically coherent lead. Lcn972-family peptides are well-characterized: they bind lipid II at the division septum and block septum invagination without forming pores ⁷. The agent’s reasoning tracks real biology.

The model doing the work is not new

The ProtNLM component is the piece with actual validation behind it, and it predates Antigravity by years. ProtNLM has been annotating UniProtKB/TrEMBL since 2022, now covers over 49 million entries, and is tagged with a distinct evidence code (ECO:0008006) to separate its predictions from human-curated Swiss-Prot ⁸. In an EMBL-EBI blind evaluation of 100 predictions, curators judged 79% demonstrably correct and preferred or found ProtNLM names equivalent to manual annotations in 88% of cases ⁹.

So the “scan millions of dark-matter sequences” story holds. What’s new in this announcement is the agentic wrapper — the conversational orchestration, the 3D viewer, the auto-generated wet-lab protocols — not the predictions themselves.

What’s missing from the discovery claim

For a result being staged as a step toward answering the 10-million-deaths-by-2050 AMR projection, the gaps are striking. An independent review by explainx.ai is blunt:

Technical reports for A0A0U0D591 did not provide essential metrics such as false-positive rates or direct performance comparisons against established AMP prediction tools like AMPlify or OmegAMP… until MIC data is published, the peptide is a sequence of interest rather than a proven antibiotic. ¹⁰

There is no minimum inhibitory concentration, no activity data against drug-resistant strains, no benchmark showing the agentic pipeline beats a tuned conventional one. Wet-lab validation is “underway” — which is where the real evidence will come from, and where this announcement should be re-read.

The dual-use and reliability shadows

DeepMind shipped SynthID Bio in September 2026 specifically because machine-generated protein sequences can “slip past conventional filters,” in Demis Hassabis’s own framing ¹¹. The Antigravity post does not mention synthesis-vendor screening or red-team findings on AI-designed toxins — a conspicuous omission for a platform whose explicit output is novel bioactive peptides plus a bench protocol to express them.

Separately, developer-forum reviewers of Antigravity 2.0 describe forced updates that overwrote legacy IDE directories and an agent-first UI that strips warning indicators once the model judges code “safe for execution” ¹². For a tool meant to drive scientific workflows that live or die on auditability, that default is backwards.

Treat this release as a capability demo of a well-integrated stack — not a drug, and not yet a discovery.

---

FLUX.3 Klein’s open-weights pitch hinges on a 28GB upsampler

Source: huggingface-blog · published 2026-10-15

TL;DR

• Black Forest Labs’ FLUX.3 Klein lands in Diffusers with a 5.6B base and JSON bounding-box editing.
• The “runs lean” story breaks: the mandatory prompt upsampler adds ~28GB VRAM on top of 23.8 GiB.
• Independent tests put Klein at 24/32 instruction checks, vs. GPT Image 2.5’s 31.
• Klein’s real moat is near-zero drift across dozens of bbox edits — a deterministic layout primitive.

What Klein actually ships

The Diffusers integration is a clean piece of engineering. FLUX.3 Image Klein is a latent flow-matching transformer that packs 2×2 VAE patches into 128-channel tokens (one token ≈ a 16×16 pixel region), stacks eight hidden layers of a Qwen3-VL text encoder into a 20,480-dim feature vector per text token, and runs five modality-independent blocks before 28 joint blocks where text, image, and reference streams attend to each other. A 4D RoPE over (t, h, w, l) is what makes the headline feature — JSON bounding boxes with src_bbox/tgt_bbox mappings on a 0–1000 grid — behave deterministically.

That determinism is the one claim that survives independent scrutiny. TechTimes’ iterative-editing test clocked Klein’s drift score near zero across dozens of consecutive edits, where competing editors degraded visibly after a single pass ¹³. For agent pipelines that chain edits — “move this object, then recolor that region, then add text here” — a numerically addressable canvas is a real primitive.

The prompt-upsampler asterisk

Everything else needs an asterisk, and the asterisk is spelled PUP. BFL’s Stephen Batifol posted on the Hugging Face discussion board that “Prompt Upsampling should be used all the time” while apologizing for a broken ComfyUI workflow at launch ¹⁴. Kingy AI’s benchmark confirmed why: with PUP enabled, hosted Klein passed 24 of 32 scored instruction checks; GPT Image 2.5 passed 31, and disabling PUP exposes weak native instruction following ¹⁵. OrcaRouter’s teardown adds the kicker — the “mandatory 9B prompt-upsampler VLM adds roughly 28GB to the total VRAM requirement,” which they call a “load-bearing dependency” that undermines the open-weights framing ¹⁶. The blog’s 23.8 GiB peak on an H100 only holds if you pretend the thing that makes Klein competitive isn’t loaded.

How it stacks up

Against Qwen-Image 2.1 Turbo specifically, Klein loses on compositional capability, native transparency, and mask handling; it wins on licensing terms and an NVFP4 Blackwell build that fits a prosumer card ¹⁷.

The license nobody is reading closely

The FLUX Kommunity License is the other quiet constraint. Free commercial use is capped at individuals and businesses under $5M annual revenue; anyone larger must negotiate a separate agreement or use BFL’s paid API ¹⁸. Combined with the PUP dependency, that makes “open weights” more of a distribution strategy than a stance — fine for indie builders, a conversation for anyone with a finance team.

The takeaway

Klein is worth integrating if you need bbox-grounded, drift-free iterative editing in a pipeline you control. It is not the open-weights frontier model the launch post implies, and the 28GB PUP is the number BFL should have put next to the 23.8 GiB.

Round-ups

Nvidia open-sources 2.8T-token Indic pretraining pipeline

Source: huggingface-blog

The Indic dataset spans 2.8 trillion tokens across major Indian languages, paired with an open data generation pipeline that others can rerun. Nvidia’s release targets the long-standing shortage of high-quality pretraining corpora for low-resource languages beyond English and Chinese.

IBM’s Granite Switch 4.1 unifies chat, decisions, and speech

Source: huggingface-blog

Granite Switch 4.1 routes chat, structured decision-making, and speech tasks through a single endpoint, cutting the usual multi-model orchestration overhead. IBM pitches the unified interface at enterprise teams juggling separate inference stacks for text, classification, and audio workloads.

Simon Willison ships two-phone Pong that networks over audio chirps

Source: simon-willison

The static HTML game, vibe-coded with Claude Opus 5.5 in Claude Code, synchronizes two phones placed near each other by encoding game state as microphone-readable chirps. No server, no Bluetooth pairing — just sound as the transport layer between browsers.

Footnotes

1. eMasterLabs review — https://emasterlabs.com/meta-muse-spark-review

   On the Artificial Analysis Intelligence Index, Muse Spark Open scores 39, trailing Kimi K3 (44) and MiMo-V2.6-Pro (46); DeepSeek V4.1 Flash leads on Terminal-Bench 4.0 by 27% vs 8%.

   ↩

2. Hacker News discussion (news.ycombinator.com) — https://news.ycombinator.com/item?id=48846184

   Meta allegedly used a harness with 6 CPU cores and 8GB of RAM for tasks that officially allow significantly less… the headline numbers are essentially ‘cheating’.

   ↩

3. The News (thenews.com.pk) — ‘critics call it overhyped’ — https://www.thenews.com.pk/latest/1398286-meta-launches-muse-spark-ai-model-critics-call-it-overhyped

   Alexandr Wang admitted the model currently fails to meet basic requirements for the ARC AGI 2 benchmark… Chollet criticized the model for appearing over-optimized for public benchmarks at the expense of practical application.

   ↩

4. Medium — ‘I tested Muse Spark for a week’ (pixipace) — https://medium.com/@pixipace/i-tested-metas-muse-spark-for-a-week-here-s-what-nobody-s-saying-44e1af41f03b

   Testers reported the model fumbles simple tasks like Mac swipe gestures or Three.js physics… DeepSeek v4.1 often completes the same tasks in a single pass where Muse Spark fails. The ‘Contributor tier’ data-sharing arrangement is ‘creepy’.

   ↩ ↩²

5. SQ Magazine — Muse Model License analysis — https://sqmagazine.co.uk/meta-muse-spark-open-weight-model-license/

   Entities with over $20 million in annual gross revenue must notify Meta and enter into a separate agreement before using the model commercially — a ‘revenue-gating’ clause critics say fails OSI Freedom 0.

   ↩ ↩² ↩³

6. e-acc.ai — NVFP4 deployment notes — https://e-acc.ai/

   SGLang silently drops weights for specific layers when loading llm-compressor NVFP4 checkpoints; vLLM’s —kv-cache-dtype nvfp4 flag triggers EngineDeadError at first inference on SM120 because TensorRT-LLM kernels are missing.

   ↩ ↩²

7. MDPI Microorganisms review on Lcn972 — https://www.mdpi.com/2076-2607/11/2/501

   Lactococcin 972-family bacteriocins are non-pore-forming and inhibit cell wall biosynthesis by specifically binding to lipid II at the division septum, blocking septum invagination without disrupting the plasma membrane.

   ↩

8. UniProt ProtNLM help page — https://www.uniprot.org/help/ProtNLM

   ProtNLM predictions are tagged with the evidence code ECO:0008006 (Google:ProtNLM) to distinguish them from human-curated Swiss-Prot entries, and have annotated over 49 million entries in UniProtKB/TrEMBL.

   ↩

9. EMBL-EBI technology news — https://www.ebi.ac.uk/about/news/technology-and-innovation/natural-language-processing-for-rapid-protein-annotation/

   In a blind evaluation of 100 random predictions, curators reached consensus that 79% were demonstrably correct, while 18% could neither be verified nor invalidated; in 88% of cases curators either preferred the ProtNLM-generated names or found them equivalent to existing manual annotations.

   ↩

10. explainx.ai analysis — https://explainx.ai/blog/deepmind-antigravity-antimicrobial-peptide-bacteriocin-s-pseudopneumoniae-agentic-science-2026

    Technical reports for A0A0U0D591 did not provide essential metrics such as false-positive rates or direct performance comparisons against established AMP prediction tools like AMPlify or OmegAMP… until MIC data is published, the peptide is a sequence of interest rather than a proven antibiotic.

    ↩

11. The Primary (AI biosecurity report) — https://theprimary.com/ai-tech/2026-09-03/ai-biosecurity-pathogen-design

    DeepMind introduced SynthID Bio in September 2026, embedding imperceptible watermarks into synthetic protein sequences… Hassabis categorized biosecurity as an ‘urgent challenge,’ warning that machine-generated sequences may increasingly slip past conventional filters.

    ↩

12. Google AI Developer Forum review of Antigravity 2.0 — https://discuss.ai.google.dev/t/google-antigravity-2-0-review-between-memory-efficiency-breakthroughs-and-migration-method-chaos/147256

    The forced direct-update method caused installation directory conflicts where the new platform and legacy IDE versions would overwrite one another, with the interface stripping away visual indicators for warnings or errors, leaving users ‘blind’ to hidden bugs if the AI deems the code ‘safe for execution’.

    ↩

13. TechTimes on bounding-box JSON control — https://www.techtimes.com/articles/328502/20261002/black-forest-labs-launches-flux-3-image-json-bounding-boxes-lock-unchanged-pixels-numerically.htm

    FLUX.3 maintained a drift score near 0 across dozens of consecutive edits using its 0–1000 normalized bbox grid, whereas competing iterative editors showed significant quality degradation after a single generation.

    ↩

14. Stephen Batifol (BFL), HF discussion — https://huggingface.co/black-forest-labs/flux-3-image-klein-comfy/discussions/8

    Hey everyone, Stephen from BFL here—we noticed there is an error in the Comfy workflow, we are working on a fix… Prompt Upsampling should be used all the time.

    ↩

15. Kingy AI benchmark review — https://kingy.ai/blog/flux-3-image-klein-tested-benchmarks-costs/

    A hosted ‘distilled’ Klein model passed 24 out of 32 scored instruction checks; GPT Image 2.5 passed 31. BFL recommends using a prompt upsampler to convert simple requests into detailed scene captions, as disabling this feature reportedly highlights weaknesses in the model’s native instruction following.

    ↩ ↩² ↩³

16. OrcaRouter on the Prompt Upsampler (PUP) — https://www.orcarouter.ai/blog/flux-3-image-klein-pup-quiet-release

    The mandatory 9B prompt-upsampler VLM adds roughly 28GB to the total VRAM requirement on 80GB GPUs, making it a ‘load-bearing’ dependency that open-source advocates say undermines the ‘open weights’ claim.

    ↩

17. OrcaRouter — Klein vs Qwen-Image 2.1 Turbo — https://www.orcarouter.ai/blog/flux-3-image-klein-vs-qwen-image-2-1-turbo

    Qwen-Image 2.1 offers native transparency and mask handling that Klein lacks, and tops open-weight GenEval categories including layout and text rendering; Klein’s edge is the Kommunity license and 8GB-VRAM NVFP4/INT8 builds.

    ↩ ↩² ↩³ ↩⁴

18. VentureBeat on FLUX Kommunity License — https://venturebeat.com/technology/black-forest-labs-opens-flux-3-image-klein-weights-with-free-commercial-use-limited-to-smaller-businesses

    Free commercial use is limited to individuals and businesses with annual revenue under $5 million; organizations above that threshold must negotiate separate commercial agreements for self-hosting or use BFL’s paid API.

    ↩ ↩²