Open Weights Approach the Frontier: 501B Beam and 816B Meta MoE Debut in the Same Week

Over the past week (2026-10-10 to 10-17), a phenomenon emerged in the large model arena that deserves its own discussion: the capability gap between closed-source flagships and open weights has been compressed to an almost negligible level. Google pushed Gemini 4 Argon to general availability, Meta released the 816B Muse Spark Open, and new lab Reflection made its debut with the 501B Beam. All three models landed in quick succession, and all of them put "agentic coding" and "million-token context" in their marketing copy.

Three heavyweight releases carry the week

Gemini 4 Argon: a pricing window for the closed-source flagship

Google announced general availability of Gemini 4 Argon on October 13, positioned for complex multi-step tasks—coding, finance, legal, cybersecurity, and scientific research—with a unified input/output context of 1 million tokens.

What developers should really pay attention to is the pricing cadence. Introductory API pricing is $2 per million input tokens and $10 per million output tokens, with a 95% discount on cached input; but this price only lasts until November 30, after which it doubles to $4 and $20. In a plugin changelog, Simon Willison compared this price adjustment with Claude Opus 5.5's pricing, and the conclusion is straightforward: if you are cost-sensitive, now is the window to use it. The model is already live in the Gemini app, AI Studio, Android Studio, Antigravity, and Gemini for Business.

What the "cached input" discount means: In input token billing for large model APIs, if multiple requests carry the same prefix (such as a long system prompt, code repository context, or multi-turn conversation history), the provider can retain and reuse the KV cache for that prefix instead of recomputing it each time. The hit portion is billed at a price far below standard input, typically 5%–25% of the original price. For agentic coding scenarios that repeatedly query the same long document or the same set of tool definitions, this discount can significantly lower actual costs. Therefore, "cached input has a 95% discount" in the article means the cached-hit portion is billed at only about 5% of the standard input price, a variable that must be included when evaluating real usage costs.

Muse Spark Open: Meta's big open-source multimodal bet

This time Meta did not announce the model on its own ai.meta.com blog (whose latest post is still from July), but instead released Muse Spark Open through Hugging Face. It is a multimodal MoE with about 816B total parameters and about 42B activated per token, with a maximum context of 1 million tokens, which Meta calls "the strongest open-weight model from the US to date."

The input side supports text, images, audio, and video, with text-only output. BF16 weights are about 1.5TB, and quantized versions can run on high-memory Macs; official support covers llama.cpp, vLLM, SGLang, MLX, and HF Inference Endpoints. In the same week, Meta also exposed the closed-source Muse Wonder and Muse Wonder Full through Simon Willison's llm-meta-ai 0.2 plugin—Willison speculates these may be Meta's first "Fable class" models, and Artificial Analysis's assessment is that the Max version is roughly on par with GPT-6.1 Sol and offers six levels of reasoning intensity.

MoE's "total parameters" vs. "activated parameters": A Mixture-of-Experts (MoE) model splits the feed-forward layers into multiple expert subnetworks, and each token is routed to only a few of them, so the number of parameters actually involved in computation during inference is far smaller than the total number of weight parameters. In the article, 816B total parameters and about 42B activated parameters mean that storage still needs to hold all ~1.5TB of BF16 weights, but the per-token compute cost is close to that of a 42B dense model. This explains why it can run as a quantized version on high-memory Macs: the bottleneck is VRAM/memory capacity rather than compute, and quantization only needs to compress the weight size to an acceptable range. When evaluating MoE, total parameters determine the deployment threshold, while activated parameters determine inference speed and compute cost; the two must be considered separately.

Beam: a new lab's 501B debut

Reflection, a new US lab, opened with frontier scale: Beam is a 501B-parameter open MoE with 23B activated per token, an Apache 2.0 license, 1 million token context, and the same target scenarios of coding and agentic tasks.

The engineering rollout is fairly complete: transformers, vLLM, and SGLang all provide Day-0 support, along with SFT/RL recipes. An NVFP4 checkpoint of about 280GB can run on 4 B200s, while full BF16 requires about 1TB of VRAM. For teams that want to own and customize a frontier model themselves, this combination of license and hardware threshold is quite attractive.

NVFP4 checkpoints and the B200 deployment threshold: NVFP4 is a 4-bit floating-point weight format that Nvidia is pushing on the Blackwell architecture. Compared with BF16, it can compress model size to about one quarter while retaining acceptable accuracy, and it can gain native acceleration at the hardware level. In the article, an NVFP4 checkpoint of about 280GB can run on 4 B200s, while full BF16 requires about 1TB of VRAM; the difference comes precisely from this compression ratio. A single B200 has around 180GB of VRAM, so 4 cards total about 720GB, enough to hold 280GB of weights while leaving room for KV cache and activations; running BF16 would require 8 or more machines of the same class. For teams building their own infrastructure, the choice of quantization format directly determines the number of GPUs needed and rental costs.

Beyond the giants: small models solving concrete problems

Another thread this week is that task-specialized small models are shipping densely, and they are often closer to real workflows than large models.

Mistral launched Timestral, its first time-series foundation model family, in four sizes (5M, 28M, 123M, and 209M parameters), for probabilistic multivariate forecasting covering sensor monitoring, anomaly detection, and capacity planning. Mistral says the 209M version ranks first among zero-shot pretrained models on GIFT-Eval/TIME/fev-bench, runs purely on CPU, takes 10–140ms per forecast, supports a 32,768-step context, uses a "Modified MIT" license, and has weights on GitHub and Hugging Face. Teams that do not want to set up GPUs for forecasting tasks can use it directly.

LightOn's ColGrep-Agent takes another path: a 2B-parameter local code search model that lets coding agents outsource repository retrieval instead of forcing grep. Tested on Claude Code, it reduced API tokens by 34% and cache-aware cost by 40% with no obvious loss in accuracy, at the cost of latency rising from 13.0 seconds to a median of 21.6 seconds. GGUF weights are already available.

IBM's Granite Switch 4.1 (8B, 30B) packages a base chat model, a Decision adapter, a Guardian safety adapter, and an embedded speech-to-text model into a single endpoint. The key is that the adapters reuse the base model's KV cache, so the VRAM overhead of the whole combination is about the same as running the base model alone, eliminating three separate deployments.

Other items worth noting: Black Forest Labs' FLUX.3 Image Klein received first-day Diffusers support, with the distilled checkpoint generating 1024x1024 images in about 2 seconds on an H100; FrontiersMind's Lumma-Fev-27B leads TypeSafe Jev on decision accuracy at 90.7% vs. 83.9%, but still trails on general reasoning; Ai2's HiRO-ACEv2 generates global 3km climate data at about 300x lower cost than X-SHIELD; DMIS Lab's UltraBERT uses a single encoder for both classification and retrieval; and Liquid AI's LFM2.5-24B-A2B activates only 2B parameters per token and reaches 184 tokens/s on an Apple M5 Max.

How "interface-compatible adapters" (LoRA-style composition) save VRAM: IBM Granite Switch 4.1 packages chat, decision-making, safety guardrails, and speech transcription into a single endpoint by having adapters reuse the base model's KV cache. Such adapters are usually low-rank incremental weights attached to the same base, sharing the base's prefix cache and attention state. Therefore, when running multiple tasks in parallel, VRAM usage is roughly "one base + several lightweight increments" rather than "multiple full models." The tradeoff is that all tasks share the same base capacity, so the capability ceiling is constrained by the base, and task switching must be scheduled within the same serving process. For teams that need to provide chat, compliance filtering, and specialized judgments from a single endpoint, this packaging approach can eliminate the operations and VRAM overhead of multiple separate deployments.

The line between open and closed source is fading

Looking at these news items together, the clearest signal is that "frontier capability" is no longer an exclusive label for closed-source models. Google holds the commercial side with a token-billed closed-source flagship, while Meta and Reflection released open-weight models of 816B and 501B within 24 hours of each other, with parameter scale, context length, and target scenarios almost aligned.

Another trend is the maturation of model composition strategies. Teams are no longer waiting for a single all-purpose model; instead, they pair giant general-purpose models with narrow-domain, low-cost small models—Timestral for forecasting, ColGrep-Agent for retrieval, Granite Switch for multi-task endpoints. This "large model as the foundation, small models doing the work" architecture may determine real-world deployment results more than simply chasing parameter scale.

As for Anthropic, there were no model-related moves this week; public activity centered on policy and business expansion (independent evaluator updates, midterm election preparation, and the Singapore office), which can be seen as a quiet observation window.

One-line Summary

Open weights and closed-source flagships are nearly at capability parity, and the composition strategy of large models as the foundation with small models doing the work is becoming mainstream.

Paragraph Summary

This week (2026-10-10 to 10-17), the clearest signal in the large model arena is that the capability gap between closed-source flagships and open weights has been compressed to near-negligible levels. Google's Gemini 4 Argon is generally available with 1 million token context and introductory pricing of $2/$10 per million tokens, but it doubles after November 30, creating a clear usage window. Meta released the multimodal MoE Muse Spark Open through Hugging Face with 816B total parameters and about 42B activated, while Reflection debuted the Apache 2.0 501B Beam, whose NVFP4 checkpoint can run on 4 B200s. Beyond the giants, task-specialized small models such as Mistral Timestral, LightOn ColGrep-Agent, and IBM Granite Switch are shipping densely, focused on low-cost deployment. The overall trend is that frontier capability is no longer exclusive to closed source, and teams are shifting to a composition architecture with a giant general-purpose model as the foundation and narrow-domain small models doing the work.