Reflection releases Beam's 501B open weights on Hugging Face under Apache 2.0

What Reflection released on 16 October

Reflection published Beam is now available on 16 October 2026. The post says the company has released Beam's model weights on Hugging Face under an Apache 2.0 licence, alongside a model card with final benchmarks and a Beam Safety Technical Report. It also says the launch comes with an ecosystem of more than 40 partners and integrations, which Reflection describes as an open operating system for agents, with the model as the kernel.

That closes the gap left by the company's 5 October announcement. When Reflection first introduced Beam, it said the model was still undergoing final red-teaming and evaluations and that weights, a technical report, a model card and developer artifacts would follow later in the month. AI Lookout covered that stage in Reflection previews Beam, a 501B open-weight model still under evaluation. Until today, access ran through a waitlisted API.

The main Hugging Face repository is public and not gated. When AI Lookout read the Hub API on 16 October, it listed 198 files, an Apache 2.0 licence tag, a LICENSE file containing the Apache License 2.0 text, and 188 safetensors shards. Separate repositories hold the FP8 and NVFP4 checkpoints and a DFlash2 drafter model for speculative decoding. Unsloth had already published its own copy and a GGUF conversion the same morning.

The model card: 501B total, 23B active, text only

The model card describes Beam as an instruction-tuned model that takes text in and produces text out. It has 501.3 billion total parameters and 23.4 billion active parameters, spread over 52 layers: 3 dense layers and 49 mixture-of-experts layers with 256 experts. Attention combines gated grouped-query attention with sliding-window attention. The context window is 1 million tokens, the vocabulary is about 152,000 tokens, the chat template is a modified Harmony format, and the knowledge cutoff is 30 June 2026. It was developed predominantly for English.

Reflection lists four primary capabilities: token-efficient reasoning, software engineering, agentic tool use, and maths and science reasoning. The efficiency claim is specific. The card says Beam reaches solution quality comparable to leading open models using about three times fewer inference-time reasoning tokens. The 5 October post put the training scale at 23.8 trillion pretraining tokens and a reinforcement-learning run of more than 100 million rollouts on 10,500 NVIDIA GB300 GPUs over four weeks.

The usage section is permissive in plain language. Reflection says it does not prohibit distillation, using outputs to train other models, fine-tuning, parameter-efficient adaptation, quantisation, merging, defensive security research on systems the user is authorised to test, or good-faith adversarial evaluation. It points to a separate usage policy as recommended reading and says the model card does not modify the licence.

Final scores moved a little from the preview

The final card reports results for all three numeric formats, run at temperature 0.7 and top-p 0.99. For the BF16 checkpoint, Reflection reports 81.6% on SWE-bench Verified, 78.3% on SWE-bench Multilingual, 66.1% on SWE Bench Pro v1, 76.5% on SWE Bench Pro v2-Hard and 80.1% on Terminal Bench v2.1. On reasoning it lists 97.7% on AIME 2026, 90.6% on GPQA Diamond, 36.0% on Humanity's Last Exam without tools and 51.1% with tools. On tool use it reports 80.8% on MCP Atlas and 76.8% on BrowseComp with context management.

Those numbers are close to, but not the same as, the preview table published on 5 October and updated on 8 October. SWE-bench Verified rose from 80.9% to 81.6% and MCP Atlas from 78.7% to 80.8%. Several others slipped: SWE Atlas Codebase QnA fell from 34.6% to 31.8%, tau3 Banking from 38.0% to 36.1%, and the AA Omniscience index from 13.0 to 10.8. Terminal Bench v2.1 stayed at 80.1%. Small moves like these are normal between a preview and a final checkpoint, but they are a reason to cite the model card rather than earlier launch graphics.

The quantised versions track BF16 closely on Reflection's own runs. NVFP4 scores 80.7% on SWE-bench Verified and 78.4% on Terminal Bench v2.1, and on SWE Bench Pro v2-Hard both FP8 and NVFP4 score higher than BF16. Every one of these figures is self-reported. AI Lookout has not run Beam, and no independent evaluation of the released checkpoints was available when this article was written.

What it takes to run the weights

Only 23 billion parameters are active per token, but every expert still has to be loaded. The Hub API showed the BF16 repository at about 1,003 GB of files, the FP8 repository at about 503 GB and the NVFP4 repository at about 294 GB. The DFlash2 drafter is about 3.9 GB. Even the smallest official format needs multi-GPU data-centre memory, and Reflection's availability post does not give a minimum hardware specification.

Reflection says Beam has day-zero support in Transformers, vLLM, SGLang, Unsloth, Axolotl, TokenSpeed and NVIDIA Dynamo, and that partners helped tune serving configurations and fused kernels. Its developer docs list the API model ID as Beam-1.0-501B-A23B, with up to 128,000 output tokens and support for reasoning, tool calling and structured outputs; the models page notes that the context window may change during the beta. For most teams the realistic first step is a hosted endpoint, not a download.

Mirror CLI and the partner list

Alongside the weights, Reflection is pushing Mirror CLI, its terminal coding harness, published under Apache 2.0. The company says it trained Beam together with Mirror, which pairs a Python front end for prompts, skills and tooling with a Rust agent core. Mirror supports resumable conversations, parallel agents in separate worktrees, MCP connections and Python extensions. Reflection says Beam has also been tested with Pi, OpenCode and Hermes.

The availability post names a long list of launch platforms: Amazon Web Services, Nscale, Applied Compute, Baseten, Cloudflare, CoreWeave, Databricks, Fireworks AI, Google Cloud, Hermes, LM Studio, Microsoft Azure, Modal, Nebius, NVIDIA NIM, Ollama, OpenCode, OpenRouter, Oracle, Together AI, Vercel and Weights & Biases, with Dell and HPE for on-premises and air-gapped deployments. The post also says Reflection's own API is free to start in public beta, without giving per-token prices.

For customisation, Reflection points to fine-tuning and reinforcement-learning workflows built with NVIDIA NeMo RL and Megatron, RadixArk's Miles, Hugging Face TRL, Anyscale's SkyRL and Axolotl. It says its largest RL runs used about 1.3 billion sandboxes across training, evaluation and ablations, and that self-distillation recipes will follow in the coming days.

Safety numbers and the missing technical report

The model card's safety section reports Reflection's in-house Iris Safety Evaluations, which the company says it will open-source. Combined across harm areas, the BF16 model refused or redirected 88.9% of harmful requests and complied with 81.5% of harmless ones. That second number means roughly one in five benign prompts in the test set was refused or deflected, which matters for teams building assistants that must not over-refuse.

The availability post links a Beam Safety Technical Report, and the card links Reflection's Open Safety Framework. But the model card header still lists the general technical report as coming, so details such as the full training-data mix, ablations and the evaluation harness behind the capability table are not yet public. The card says pretraining used public web data and data obtained through commercial partner agreements, and that reinforcement learning drew on synthetic pipelines plus proprietary vendor data.

The post also thanks external partners that tested Beam's capabilities and safety before release, including Andon Labs, Artificial Analysis, Arena, Gray Swan, Irregular, SecureBio, Snorkel AI and Mercor. Reflection does not publish those partners' findings in the post itself.

Where Beam sits among this month's open models

Beam arrives in a crowded fortnight for large open models. Meta released Muse Spark Open a day earlier with 816 billion total and 42 billion active parameters. Ant Group published Ling 3.1 Flash under MIT, and NVIDIA added Nemotron 3.5 Super VL and Bolt. Reflection positions Beam as a Western open-weight option that is efficient at inference time rather than the largest model on every chart.

The preview post said as much. It described Beam as competitive with GLM 5.2 and approaching Qwen 3.8 Max on coding and agentic tasks, while acknowledging that Kimi K3 remains ahead on raw capability. The licence is the clearest differentiator: Apache 2.0 with explicit permission to distil and train on outputs is looser than many large-model licences, which often add acceptable-use addenda or user-count thresholds.

What to check before you build on it

Before committing to Beam, teams should check four things. First, confirm the licence file in the exact repository and format they download, since BF16, FP8, NVFP4 and the drafter live in separate repos. Second, size hardware against the real file sizes, not the 23 billion active parameters. Third, rerun the benchmarks that matter for their workload, because every number in the card is vendor-reported. Fourth, watch for the technical report, which should explain how the scores were produced.

The durable news is simple. The weights Reflection promised on 5 October are now public, ungated and under Apache 2.0, in three precisions, with a hosted API and a long partner list behind them. Whether Beam's claimed efficiency holds up outside Reflection's own evaluations is the next thing to measure.