Two US Open-Weight Drops in 24 Hours Can't Touch DeepSeek's V4-Pro-0813 — a 1.6T/49B MoE at $0.87/M Output Tokens

Meta shipped Muse Glimmer on Aug 11. Nvidia shipped Nemotron 3.5 Lightning on Aug 12. Same window, DeepSeek shipped V4-Pro-0813 — a 1.6T-total/49B-active MoE at $0.87/M output tokens. Two consumer-GPU-friendly 30B MoEs vs. one frontier MoE. The US open-weight response is real but a tier behind.

DeepSeek’s V4-Pro-0813 hit GA yesterday — a 1.6T-total / 49B-active Mixture-of-Experts under MIT license, priced at $0.87 per million output tokens with $0.435/M cache-miss input and $0.003625/M cache-hit input. In the same 24-hour window, Meta shipped Muse Glimmer (30B, open weight) on Aug 11 and Nvidia shipped Nemotron 3.5 Lightning (30B open-weight MoE) on Aug 12. Two American 30Bs you can fit on a consumer GPU went up against one Chinese 1.6T frontier model. The US open-weight response is real — and a tier behind.

The open-weight timeline, Aug 11–12

Aug 11 — Meta Muse Glimmer (30B dense, Apache 2.0). Meta Superintelligence Labs’ first open-weight release in over a year, distilled from the larger Muse Spark. Multimodal, 131k context, K-Quant-17GB fits a single RTX 3090. I covered it yesterday — the DFlash drafter shipping pre-integrated is the technical headline.

Aug 12 — Nvidia Nemotron 3.5 Lightning (30B-total / 3B-active MoE, open license). Nvidia’s first explicitly open-weight model. The “Lightning” name signals throughput: positioned for always-on agents and high-concurrency serving, not single-prompt quality. Pricing isn’t tied to a hosted API because there isn’t one — Nvidia ships weights and expects you to bring your own GPU. The 30B/3B split mirrors the Qwen3 35B-A3B pattern from earlier in 2026: a big MoE where only a tiny fraction fires per token, so it runs cheap on commodity hardware.

Aug 12 — DeepSeek V4-Pro-0813 (1.6T total / 49B active, MIT). Preview since April 24; Flash went GA July 31; Pro followed Aug 12 with the 0813 build. Same hour Nemotron 3.5 was announced, DeepSeek shipped the production cut of its flagship.

The framing matters: Meta and Nvidia both shipped “the open-weight model you can run on one GPU.” DeepSeek shipped “the open-weight model that beats the closed frontier labs on benchmarks.” Different leagues.

What V4-Pro-0813 actually is

V4-Pro-0813 is a Mixture-of-Experts model with 1.6 trillion total parameters but only 49 billion active per token. That’s a 32× expert-to-active ratio — extremely sparse routing, the kind of design choice that lets you scale total capacity without scaling inference cost proportionally. Context is 1 million tokens, max output 384k. Weights are on Hugging Face under MIT.

The architectural twist is two attention variants DeepSeek ships in parallel: Compressed Sparse Attention and Heavily Compressed Sparse Attention. CSA and HCSA both reduce the KV-cache footprint aggressively — useful when 1M context is a target. The exact compression ratios and routing logic are in the model card.

Per Wccftech, DeepSeek is now #2 globally in token consumption, second only to Anthropic. That is not a hype claim — it’s a measured share of API traffic across major aggregators. The Chinese lab isn’t shipping a model and hoping. It’s shipping a model and watching traffic move.

The pricing that changes the math

Tier Input (cache miss) Input (cache hit) Output Concurrency
DeepSeek V4-Pro-0813 $0.435 / M $0.003625 / M $0.87 / M 500
DeepSeek V4-Flash — — — 2,500

Cache-hit input at $0.003625 per million tokens is the number that kills closed-frontier economics for any workload with a stable system prompt — agent loops, retrieval pipelines, anything where most of the context is reused across calls. DeepSeek-Pro’s concurrency cap is 500 simultaneous requests; Flash goes to 2,500. If you need higher throughput, Flash is the path; if you need the flagship quality at this price, Pro is the path.

For comparison: closed-frontier output pricing still sits around $15–$60 / M tokens depending on tier. V4-Pro-0813 at $0.87 / M is roughly 17–70× cheaper. The cache-hit input ratio is even more lopsided. This is the same story DeepSeek told with R1 a year and a half ago — they’re just better at it now and shipping frontier-tier quality.

The honesty caveat

DeepSeek claims +15.8% on Terminal Bench versus the April preview build. That number is vendor-reported. No third-party evaluator has replicated the V4-Pro-0813 benchmark scores yet — TechTimes flagged this on GA day. Treat the headline numbers as DeepSeek’s own measurements until Artificial Analysis, Hugging Face’s OpenLLM Leaderboard, or independent reproduction publishes a comparable score.

This matters because the price story is only interesting if the model actually delivers. If the 15.8% Terminal Bench number holds under independent testing, V4-Pro-0813 is the new default for any agent workload that doesn’t require Anthropic or OpenAI specifically. If it doesn’t, the price is still aggressive but the tier-shift framing collapses.

Why the US 30Bs matter anyway

Nvidia and Meta didn’t lose this round — they entered a different one. Nemotron 3.5 Lightning’s 3B-active design means it runs at single-GPU inference cost with a 30B-quality ceiling. Muse Glimmer’s K-Quant-17GB build is a real consumer-GPU agent model. The CNBC framing — “open-weight AI race with China” — understates it: these are two different products.

  • DeepSeek V4-Pro-0813: frontier-tier quality at open-weight pricing, MIT license, runs on DeepSeek’s own infra or anyone willing to host 1.6T parameters. Target user: API customers and labs with serious GPU budgets.
  • Nvidia Nemotron 3.5 Lightning: open weights, designed for self-hosting, throughput-optimized for agent loops. Target user: enterprises running their own inference fleet.
  • Meta Muse Glimmer: open weights, multimodal, consumer-GPU fit. Target user: developers running local agents on a workstation or a 3090.

All three shipped in 24 hours. None of them competes head-to-head. But the narrative collapse — “US open weight vs Chinese open weight” — is real, and the gap is widening on the frontier tier while staying narrow on the consumer tier. A year ago you could argue the US was ahead on both. Today the open-weight frontier is Chinese, and the consumer-GPU open-weight segment is contested.

What to do with this

If you’re picking a model today:

  • Need a hosted frontier-tier open-weight model with MIT terms and aggressive pricing? V4-Pro-0813. Pin the cache-hit pricing — that’s where the savings compound.
  • Need an open-weight agent model you can self-host on one GPU? Either Nemotron 3.5 Lightning (if you’re on Nvidia hardware and want throughput) or Muse Glimmer (if you want multimodal and a speculative drafter out of the box).
  • Need both? Run Glimmer locally for dev and short calls, route long-context or batch jobs to V4-Pro-0813 via API. The price gap means the hybrid is cheaper than running either alone at scale.

The bigger signal: 1.6T/49B at $0.87/M is the new floor for frontier-tier open-weight models. Whatever Nvidia and Meta ship next has to beat that on price, license terms, or local runnability — not on raw capability.

Sources