Meta opens Muse Spark at 816B, Reflection releases Beam under Apache 2.0

Meta has opened the weights of Muse Spark Open, an MoE with 816B parameters, of which 42B are active per token. The model scored 39 on the Artificial Analysis index. That is the best result among American open-weights models, 13 points above NVIDIA Nemotron 3.5 Super VL (26). Almost simultaneously, Reflection released Beam 1.0 with 501B parameters (23B active) under Apache 2.0. Its score of 30 is second among American open models.

American open weights: power versus agentic ability

Muse Spark Open is built on 62 layers with GQA (16 KV heads). It activates 8 of 256 experts per token, has a 1M context window, and accepts text, images, video, and audio as input. The BF16 weights take up about 1.

6 TB. The NVFP4 checkpoint quantizes only the experts and shrinks the weights to ~530 GB. According to Meta, it fits on 4 Blackwell Ultra GPUs, 8 H200s, or two Mac Studio M5 Ultra machines with 512 GB each.

Together, Fireworks, and Baseten charge $1.25/$4.25 per million input/output tokens.

The cost of running the index is $0.84, versus $2.01 for GLM-5.

3 and $0.27 for DeepSeek V4.1 Flash.

The model also spends just 40k output tokens per task, the fewest of the top 9 open models.

The weak spot is agentic tasks: on AA's four agentic tests it ranks last among the top 9 open models, although it leads on AA-Omniscience. Vals ranks it 12th among open models and 38th overall. In Code Arena WebDev it has 1511 points. The Muse Model License requires the label "Built with Muse Spark Open". Critics such as @teortaxesTex call the release "DOA" next to the smaller, MIT-licensed V4.1 Flash.

Reflection Beam 1.0 ships with BF16, FP8, and NVFP4 weights, the DFlash2 drafter, the open Mirror CLI harness, and a safety report. It scores 48.5% on AutomationBench-AA but only 3.0% on Terminal-Bench 4.0. On AA-Omniscience it answers correctly in 33% of cases and hallucinates in 47%. In Code WebDev it is the only non-Chinese model in the top 25 open models.

• Step 5 Preview from StepFun: open weights, available through Together.
• Hy Memory 2.0 from Tencent: Hy-Memory-30B-A3B, which the company claims beats Opus 5 on memory frameworks.
• Indra from Datalab: a 1M-parameter OCR model that runs 15 times faster than Tesseract on CPU.

The closed frontier and "decision models"

• Gemini 4 Argon: the output limit rose from 64K to 1M tokens, and the introductory price of $2/$10 holds until November 30. The model is fifth in Agent Arena (+13.2%), and Perplexity set a new SOTA on the WANDR benchmark.
• Anthropic Haiku 5.5: $0.10/$0.50 for requests up to 100K tokens, roughly 90% cheaper than Haiku 4.5. It scores 39.2% on Terminal-Bench 4.0 versus 16.4% for GPT-6 Luna.
• Opus 5.5 fast mode: 2.5x speedup (according to unofficial data).
• GPT-6 Astra Pro in ChatGPT runs several agents in parallel on different parts of a task.
• GPT Live 1.1: better multilingual voice, with nine new presets recorded with native speakers.
• Elorian Glimpse: a VLM that uses recurrent computation over latent visual states. It is claimed to match or beat Astra on 5 of 12 benchmarks at token prices 17 times lower.
• Liquid d1-ultra: 65 points in the AA Decision Index, ahead of Jev 1.14 (57). The share of answers without hallucinations rose from 2% to 85%, while accuracy fell from 41% to 32%. Time per task is 92 ms, and input costs $0.042 per M tokens.
• Perplexity pplx-decider v1.2: open weights, a 27B model at 2¢/M and a 4B Flash model at 1¢/M, with added reranking and inference on multiple questions in a single pass (up to 5x faster).

Hardware, agents, and money

• TPUv7: SemiAnalysis estimates its perf per dollar at 1.42x that of GB300 NVL72 FP4 and 2.5x that of MI355X. At medium and high interactivity it is up to 3.16x better than GB200.
• Vera Rubin NVL72: up to 5.7x higher throughput per megawatt than GB300 on DeepSeek V4 Pro, and 2.8x on Kimi K3. The record is 906 tokens/s per user on V4.1 Flash. NVIDIA itself submitted the results.
• AMD MI355X: eight cards deliver 24.4% more throughput than HGX B300 on Kimi K3, and 87.4% more tokens per dollar.
• Modal ImageFS, Zephon on TorchTitan, and Cursor Environments API (up to 500 repositories, 5x faster startup) cover the infrastructure side.
• Claude Mythos 5 with OpenEvolve beat the best known quantum circuits in 7 of 8 tasks. Four Gemini 4 Argon agents beat Minecraft through the Ender Dragon in 27 h 14 min, dying 605 times and spending about $520.

In industry news: according to Reuters, Anthropic is discussing a valuation of about $1.5T instead of the expected $2T. CoreWeave shares fell 20% over three sessions. China is considering licensing open models above 2T parameters or 10^26 FLOPs.

What to check in practice

If you need an open model for agents, relying on the overall index is risky: Muse Spark leads on knowledge but sags on agentic tests. Beam under Apache 2.0 is easier to use commercially but is weak at terminal tasks. Before choosing, run your own set of tasks with real tools. Vera Rubin and MI355X suit high interactivity, while for mass inference it is worth comparing TPUv7 on price per token. For cheap classifiers in agent chains, Liquid d1-ultra and pplx-decider are worth testing.