Reflection releases Beam: a 501B open model with a 1M‑token context window
Beam is an Apache 2.0, self‑hostable model focused on coding and agent workflows, combining long context with a Mixture‑of‑Experts design that activates 23B parameters per token. Early third‑party analysis places it near the top of open‑weight systems in the U.S., highlighting strong agent performance and areas to watch.
The business desk's take → Microsoft and NVIDIA publish open-weight AI risk guide as new models arrive30-second briefing
- Reflection released Beam 501B open weights (23B active).
- Enables self-hosted long-context coding and agentic workloads.
- Watch hardware viability (NVFP4 on Blackwell) and peer benchmarks.
One-Line Summary
Open-weight models, tools, and methods converge on practical agent workflows: Beam opens at frontier scale with a 1M-token context, Apertus 1.5 adds multimodality and a thinking mode, and LEASH shows how to escalate to stronger models only when uncertainty demands it.
Open Source & Repos
Reflection releases Beam: a 501B open model with a 1M-token context
Beam is an open-weight text model you can self-host and customize for coding and agent tasks, combining a one-million-token context window with a sparse Mixture of Experts (MoE) design so each token is processed by a few “specialists” instead of the entire network. The model totals 501 billion parameters but activates about 23 billion per token, and it is released under the Apache 2.0 license as a large language model (LLM). 1
For deployment, weights are available in bfloat16 (BF16), 8-bit floating point (FP8), and NVIDIA’s 4‑bit floating-point format (NVFP4). The BF16 checkpoint is about 1 TB, while NVFP4 shrinks it to roughly 280 GB. Day‑0 support spans the Transformers library plus popular inference engines, and a Rust-based harness is included. Reflection recommends planning around four B200 graphics processing units (GPUs) to comfortably serve long contexts; a key–value (KV) cache design holds only nine global layers, bringing a 1M‑token cache to about 36.08 GiB per sequence in BF16. 1
Initial benchmarks report strong scores on agentic coding and reasoning: for example, AIME 2026 reaches up to 98.0% (FP8) and SWE Bench Pro v2‑Hard up to 78.6%, with long‑context tasks supported by a dedicated response template and tool‑calling grammar. The model also exposes a reasoning control (“low” to “max”) and a speculative decoding drafter for throughput. 1
Independent analysis on the Artificial Analysis Intelligence Index scores Beam at 30, placing it second among U.S. open‑weight models tested; it completes 34% on the Cyber Index and shows token‑efficient reasoning (~26k output tokens per task), while Terminal‑Bench 4.0 (3.0%) surfaces a notable weakness to watch. 2
Apertus 1.5 adds images, audio, and 262k context
Apertus 1.5 extends the project’s 8B and 70B open‑weight models with native image and audio understanding, an optional “thinking mode” for harder reasoning, and a four‑times longer context window of 262,144 tokens — all while keeping text‑only generation. The weights remain available under Apache 2.0 with open data and training details. 3
The team adds 4 trillion multimodal/text tokens to the 8B and 2 trillion to the 70B models, and introduces multimodality via early‑fusion tokenization: a 512×512 image becomes 1,024 tokens (Emu3.5 tokenizer) and audio is 40 tokens per second (WavTokenizer), so the architecture stays unchanged and training runs at text‑only speed. Post‑training spans supervised fine‑tuning (SFT), long‑context SFT, reinforcement learning with verifiable rewards (RLVR), direct preference optimization (DPO), and charter alignment (SDPO). 3
Apertus 1.5 also emphasizes transparent, reproducible evaluation: results are published in the Every Eval Ever (EEE) schema and mapped to regulatory reporting needs, with the team noting coverage of 56 of 60 elements across the EU AI Act’s Article 55 and the GPAI Code of Practice. Intermediate checkpoints and data mixtures are public for re‑runs and ablations. 3
Mastra adds agent scopes for safer multi-tenant apps
Mastra is a modern TypeScript framework for AI applications and agents; the 1.78.0 release introduces Agent Scopes to support multi‑tenant identity and safer memory/thread routing. Agents now accept scopes in the form “:” per call, via agent config, or injected by trusted middleware through the MASTRA_SCOPES_KEY, and Mastra merges all three. 4
This update targets practical isolation in production agent apps, helping route state to the right tenant by design. The release is tagged as of Oct 16, 2026, and the repository includes packages, docs, and community links for adoption. 4
Research Papers
LEASH: uncertainty-aware escalation and oversight for computer-use agents
LEASH is a training‑free framework that lets a smaller “weak” agent run by default, then escalates to a stronger model only when needed — and keeps the stronger model under oversight after handoff. It uses latent uncertainty estimated from the weak model’s hidden states, comparing the current state to past successful and failed trajectories to decide when to route control. 5
On OSWorld, LEASH improves over standalone weak agents and beats random escalation by 6.55–11.94 points at the same escalation rates. Using Qwen3.5‑35B‑A3B, it reaches 73.66% task success at $0.4117 per task, matching always‑on Claude Opus 5 (72.90%) while cutting inference cost by 19.2%. 5
Why It Matters
Frontier‑scale open weights are becoming usable building blocks: Beam pairs long context and a sparse MoE design with day‑one tooling, Apertus 1.5 couples multimodality with reproducible evaluations, and LEASH shows a path to control costs by escalating only when uncertainty is high. Together, they point toward agent systems that are both transparent and operationally efficient. 1
Terms in This Article
- Token
A token is the basic unit an LLM uses to represent and process text instead of reading raw characters or full words directly.
- Context Window
A context window is the finite working memory of a language model that encompasses all tokens it can reference during generation — including prior messages, current input, and even parts of its output — distinct from training data, and subject to accuracy and recall degradation as token counts grow, which makes careful curation of in-context content critical.
- MoE
Mixture of Experts (MoE) is a sparse conditional-computation architecture in which a gating/router network selects a small subset of expert feed-forward subnets for each input token—typically replacing the Transformer’s FFN—so total parameter capacity can grow via many experts while per-token compute remains bounded.
Two desks on this story
Reflection releases Beam: a 501B open model with a 1M‑token context window
Microsoft and NVIDIA publish open-weight AI risk guide as new models arrive
The companies publish a shared taxonomy and evaluation approach for open-weight model deployments. The guidance lands as NVIDIA ships Nemotron 3.5 Super VL and Reflection unveils Beam, expanding self-hosted options for agents and long-context work.
Comments (0)