Latent Space: Meta's 816B Open Model Tops US Charts, TPUv7 Beats GB300 on $/perf

Latent Space · rss · 2026-10-17

Latent Space's 10/15–10/16 AI news roundup, piloting a new show High Taste Testers. Key items:

Open weights

Meta Muse Spark Open: 816B-total/42B-active MoE, Artificial Analysis Intelligence Index 39 — best US open model, 13 points above NVIDIA Nemotron 3.5 Super VL. 62 layers, 1M context, omni input; BF16 weights 1.6TB (NVFP4 530GB, fits 4 Blackwell Ultra or 8 H200). Hosted at $1.25/$4.25 per M tokens; $0.84 per index task vs $2.01 for GLM-5.3 and $0.27 for DeepSeek V4.1 Flash. Weakness: last among top open models on AA's four agentic evals. Custom license requires attribution; MaaS >$20M revenue needs a separate deal. Day-0 support in mlx-lm/vLLM; Apple blogged RDMA-distributed inference across Mac Studios. Community calls it "DOA" against smaller MIT-licensed V4.1 Flash.
Reflection Beam 1.0: 501B/23B-active MoE, Apache 2.0, 1M context, with BF16/FP8/NVFP4 weights, DFlash2 drafter, Mirror CLI, safety report. AA 30 (#2 US open); 48.5% AutomationBench-AA but 3.0% Terminal-Bench 4.0, 47% hallucination on AA-Omniscience. Sliding-window attention (512, 5:1), no RoPE, QK-norm, logit softcapping. 1769 tok/s on 4×B300 NVFP4; SkyRL day-0 LoRA/RL with +62% LoRA decode throughput. Ollama free hosting for a limited time.
Also: StepFun Step 5 Preview open weights; Tencent Hy-Memory-30B-A3B claims to beat Opus 5 on memory frameworks; Datalab's 1M-param OCR model Indra runs 15× faster than Tesseract on CPU.

Frontier models

Gemini 4 Argon rolls out to Pro subscribers and AI Studio; output limit raised to 1M tokens; intro pricing $2/$10 until Nov 30; SOTA on Perplexity's WANDR, #5 in Agent Arena.
Anthropic: Haiku 5.5 at $0.10/$0.50 (≤100K prompts), 90% cheaper than Haiku 4.5; 39.2% on Terminal-Bench 4.0 vs 16.4% for GPT-6 Luna. Reported Opus 5.5 fast mode at 2.5×. Claude Code now gzips conversation payloads.
OpenAI: GPT-6 Astra Pro runs multiple parallel agents per task; GPT Live 1.1 improves multilingual voice + 9 new native-actor presets; unconfirmed rumor of new Codex models.
Elorian Glimpse: visual-thinking VLM doing recurrent computation over latent visual states; claims matching/beating Astra on 5 of 12 benchmarks at 17× lower token price.

Decision models

Liquid d1-ultra tops AA Decision Index (65), 85% non-hallucination rate, 92ms per task.
Perplexity open-sources pplx-decider v1.2 (27B @2¢/M, 4B Flash @1¢/M) with 5× faster single-pass multi-question inference. OpenAI Live API adds a Decisions API.

Inference hardware

SemiAnalysis: TPUv7 delivers 1.42× perf/$ of GB300 NVL72 FP4 on AgentX, 2.5× vs MI355X; up to 3.16× better than disaggregated GB200 at medium-high interactivity.
Vera Rubin NVL72: up to 5.7× GB300 throughput per MW on DeepSeek V4 Pro; record 906 tok/s per user on V4.1 Flash.
AMD MI355X: 8 GPUs give 24.4% more throughput than HGX B300 on Kimi K3 at 50 tok/s/user, and 87.4% more tokens per dollar.

More from Infra