Reflection has released Beam, a 501-billion-parameter open-weight AI model the company says was built to handle coding, reasoning, and agentic workloads at a fraction of the inference cost of comparable systems. The announcement, confirmed by multiple outlets including TechCrunch and Fortune, positions Beam as one of the largest open-weight models to come out of a Western AI lab this year, and the company is framing it explicitly as a bid to keep pace with China’s fast-moving open-model ecosystem.
Beam is a sparse Mixture-of-Experts (MoE) model: 501 billion total parameters, but only 23 billion of them activate per token. That gap between total size and active compute is the whole point of the release, and it is why Reflection is pitching Beam less as a raw scale play and more as an efficiency play aimed squarely at developers who need large-model reasoning without large-model billing.
Don't miss new tech stories on Google
Add Tech Insider once in the Google app and our stories appear in your news suggestions.
What Is Beam? Inside Reflection’s New Open-Weight Model
Beam is a text-only model, according to reporting from TechCrunch and Fortune, meaning it does not process images, audio, or video. That is a deliberate scope decision: rather than building a multimodal generalist, Reflection built a model tuned for three use cases, coding, reasoning, and agentic tasks, the kind of long-horizon, multi-step workflows that AI coding assistants and autonomous agents rely on.
The model is released under an Apache 2.0 license, one of the most permissive open-source licenses available, which allows commercial use, modification, and redistribution without the kind of usage restrictions attached to some other “open” AI releases. Reflection has described Beam as a model built for the “Western open ecosystem,” and the Apache 2.0 choice reinforces that positioning: it lowers the barrier for US and European startups to build on top of Beam without licensing friction.
As of the announcement, Beam was not yet fully public. Reflection said the model was undergoing final red-teaming and evaluations, with early access available only through a waitlist. The company said weights, a technical report, a model card, and developer artifacts were planned for release later in the month, a staged rollout pattern that has become standard for major open-weight launches as labs try to catch safety issues before the weights are downloadable by anyone.
Beam’s Architecture: 52 Layers, 256 Experts, and a Sparse MoE Design
According to specifications reported alongside the launch, Beam runs 52 layers and, within each Mixture-of-Experts layer, maintains a pool of 256 experts. For any given token, the routing mechanism activates just 8 of those 256 experts. That routed-expert count, 8 out of 256 per token, is the architectural detail that explains the 23-billion active-parameter figure: the model carries enormous total capacity but only ever “wakes up” a slice of it for each piece of text it processes.
This is not a new idea in principle. Sparse MoE architectures have been the backbone of several high-profile open releases over the past two years, precisely because they let model builders chase larger total parameter counts without a proportional increase in serving cost. What makes Beam notable is the scale at which Reflection applied the pattern: 501 billion total parameters is a large number even in an MoE context, putting Beam in the same size class as some of the biggest open releases from any lab, Western or Chinese, this year.
Why Only 23 Billion Active Parameters Matters for Inference Cost
Inference cost for large language models scales much more closely with active parameters than with total parameters, because active parameters determine how much compute actually runs for each token generated. A dense 500-billion-parameter model would require activating essentially all of its weights for every token, a punishing compute bill at scale. Beam’s sparse design means a request routes through roughly 23 billion parameters’ worth of compute instead, closer to the cost profile of a mid-sized dense model while, in theory, retaining access to the broader knowledge and specialization spread across the full 501-billion-parameter pool.
That efficiency argument is central to how Reflection is marketing Beam. The company said Beam uses “3-4x less inference compute” than leading Chinese open models like GLM-5.2, a company claim that, as of the announcement, had not been independently verified by third-party benchmarking. Reflection framed the achievement as a product of high-compute reinforcement learning applied during training, stating that this approach let the team make Beam “extremely efficient at reasoning, delivering competitive performance on coding and agentic tasks at a fraction of the token cost and inference time compute.”
The 1 Million Token Context Window Explained
Beam ships with a 1 million token context window, according to the specifications reported around the launch. For reference, a context window of that size is roughly large enough to hold several long novels, or a mid-sized codebase with its dependency tree, in a single prompt. For the coding and agentic workloads Reflection says it built Beam for, that matters practically: agents that need to reason across a large repository, a long conversation history, or a stack of retrieved documents run into context limits fast on smaller windows, and a 1 million token ceiling gives developers a lot more room before they have to start chunking or summarizing inputs.
Large context windows have become a competitive battleground among open-weight labs this year. Apertus 1.5, the Swiss AI initiative’s fully open model, launched with a 262,000-token window, well below Beam’s figure, underscoring how fast the ceiling has moved even within a single release cycle.
Pretraining Scale: 23.8 Trillion Tokens Behind Beam
Reflection trained Beam on 23.8 trillion tokens of pretraining data, according to the specifications circulated with the announcement. That figure places Beam’s training corpus in the same general range as the largest publicly disclosed pretraining runs from major open-weight labs over the past year, though direct comparisons are difficult because labs vary widely in how they define and report token counts, deduplication, and data mixture weighting.
What the figure does confirm is that Beam is not a lightly trained checkpoint repackaged for a press cycle. A near-24-trillion-token run represents a substantial compute investment before any reinforcement learning or fine-tuning stage, consistent with Reflection’s positioning of Beam as a frontier-class release rather than a smaller efficiency demo.
Apache 2.0 License: What Developers Can (and Can’t) Do With Beam
The Apache 2.0 license Reflection attached to Beam is one of the more business-friendly options in open-source AI. It permits commercial use, modification, and redistribution, and it does not require derivative works to be released under the same license, unlike copyleft alternatives. For enterprises evaluating open-weight models, license terms are often as decisive as raw benchmark scores, since legal review of a restrictive or ambiguous license can stall adoption for months.
Once Beam’s weights become publicly downloadable, the typical pattern for loading an Apache-licensed Mixture-of-Experts checkpoint through the Hugging Face ecosystem looks like this:
from transformers import AutoModelForCausalLM, AutoTokenizer
model_id = "reflection-ai/beam" # placeholder; confirm exact repo ID at release
tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForCausalLM.from_pretrained(
model_id,
device_map="auto",
torch_dtype="auto"
)
inputs = tokenizer("Explain the routing logic in a sparse MoE model.", return_tensors="pt")
output = model.generate(**inputs, max_new_tokens=256)
print(tokenizer.decode(output[0], skip_special_tokens=True))That snippet is illustrative of the standard loading pattern developers use for open-weight checkpoints, not an official Reflection code sample; the company’s own documentation and model card, once published, will be the authoritative reference for the exact repository ID and recommended serving configuration.
Early Access, Red-Teaming, and the Release Timeline
Reflection’s announcement described Beam as still undergoing final red-teaming and evaluations at launch, with access limited to a waitlist rather than an immediate public download. That is a notable contrast to some open-weight releases that drop weights the same day as the announcement blog post. Reflection instead said the full package, model weights, a technical report, a model card, and developer artifacts, would follow later in the announcement month.
The staged approach gives Reflection a window to incorporate red-teaming findings before the weights are irreversible in the wild, a meaningful distinction for a model explicitly built for agentic tasks, where a model acting autonomously on a user’s behalf carries more risk than a chat-only assistant. It also means that, as of this writing, independent researchers have not yet had the chance to run their own evaluations against the model directly, since the earliest access has gone only to waitlisted users.
Reflection’s Efficiency Claim: 3-4x Less Inference Compute
The single most aggressive claim in Reflection’s announcement is that Beam uses three to four times less inference compute than GLM-5.2, a leading Chinese open model, while remaining competitive on coding and agentic benchmarks. That is the kind of claim that, if independently verified, would reorder how enterprises think about serving costs for large reasoning models. It is also, as of the announcement, an unverified company claim rather than a peer-reviewed or third-party-benchmarked result.
Reflection’s own framing leans on reinforcement learning as the mechanism behind the gain. The company said high-compute reinforcement learning let it make Beam efficient specifically at reasoning, rather than relying purely on parameter scale to drive quality. That is consistent with a broader shift across the industry over the past year, where labs have increasingly used RL-based post-training to squeeze more reasoning performance out of a given parameter budget rather than simply training larger base models.
What Reflection Says vs. What’s Verified
It is worth separating what Reflection has stated from what independent parties have confirmed. Confirmed, by multiple outlets reporting on the same announcement: the 501-billion total parameter count, the 23-billion active parameter count, the Apache 2.0 license, the 1-million-token context window, the architecture specifics (52 layers, 256 experts, 8 routed per token), and the 23.8-trillion-token pretraining run. Unverified, because it comes solely from Reflection’s own announcement and has not yet been tested by outside researchers: the 3-4x inference compute savings claim, and the implicit comparison that Beam “advances the frontier for the Western open ecosystem.”
Competitive Comparison: Beam vs. the Open-Weight Field
Beam does not land in an empty field. Western and Chinese labs have both pushed major open-weight releases through 2026, and Reflection’s framing, explicitly positioning Beam against “Chinese models” on cost efficiency, only makes sense against that backdrop. DeepSeek’s V4.1 Flash-1017 release has been one of the most closely watched Chinese open models this cycle, while Western labs have answered with releases like Mistral’s Large 4, which launched at 1.05 trillion total parameters, and Meta’s Muse Spark Open, an 816-billion-parameter release reportedly capable of running on two Mac Studio systems.
| Model | Total Parameters | Context Window | License / Access | Notable Detail |
|---|---|---|---|---|
| Reflection Beam | 501B (23B active) | 1,000,000 tokens | Apache 2.0 | Early access via waitlist; weights pending full release |
| Mistral Large 4 | 1.05T | Not disclosed in launch specs | Commercial license | Priced at $1.36 per million tokens at launch |
| Meta Muse Spark Open | 816B | Not disclosed in launch specs | Open weights | Reported to run on two Mac Studio systems |
| Nvidia Nemotron 3.5 Super VL | 120B | Not disclosed in launch specs | Open weights | Positioned by Nvidia as roughly 5x faster than the prior generation |
| Apertus 1.5 | Not disclosed in launch specs | 262,000 tokens | Fully open (Swiss AI initiative) | Built around transparency and open governance |
| DeepSeek V4.1 Flash-1017 | Not disclosed in launch specs | Not disclosed in launch specs | Open weights | Reported benchmark score of 77.0; opened a search API |
What stands out in that lineup is less any single spec and more the pace: six major open-weight releases, spanning three continents’ worth of labs, within a matter of weeks. That is the environment Beam is launching into, and it is the backdrop against which Reflection’s efficiency claim will eventually get tested by independent benchmarkers, developers running their own cost comparisons, and rival labs with an incentive to scrutinize the numbers closely.
Beam’s Full Architecture Spec Sheet
| Specification | Detail |
|---|---|
| Total parameters | 501 billion |
| Active parameters per token | 23 billion |
| Architecture type | Sparse Mixture-of-Experts (MoE) |
| Layers | 52 |
| Experts per MoE layer | 256 |
| Routed experts activated per token | 8 |
| Context window | 1,000,000 tokens |
| Pretraining data volume | 23.8 trillion tokens |
| Input modality | Text-only |
| License | Apache 2.0 |
| Primary target workloads | Coding, reasoning, agentic tasks |
Historical Context: From Llama to the Open-Weight Surge of 2026
Open-weight AI releases have gone from a relative novelty to a near-weekly occurrence over the past two years. What started as a handful of labs publishing smaller checkpoints has turned into a sustained competitive race, with total parameter counts climbing into the hundreds of billions and, now with Mistral Large 4, past a trillion. Reflection’s decision to frame Beam explicitly as a response to Chinese open models reflects a narrative that has hardened across the Western AI industry through 2026: that open weights are no longer just a developer-goodwill gesture, but a geopolitical and commercial battleground.
Inference efficiency has become the newer front in that battle. Where the earlier phase of the open-weight race was largely a parameter-count contest, labs are now competing on tokens-per-dollar and tokens-per-watt almost as much as on raw benchmark scores, because enterprise buyers evaluating agentic deployments care directly about the recurring cost of running a model in production, not just its one-time training bill. Beam’s 23-billion active-parameter design, carried inside a 501-billion-parameter shell, is Reflection’s entry in that efficiency contest, and it mirrors a design pattern that rival labs, including the teams behind Grok, Claude, and DeepSeek’s competing model lines, have leaned on with their own sparse architectures.
Market Impact: Why a 501B Open Model Moves the AI Industry
A release at Beam’s scale matters to more than just the open-source AI community. Enterprises building AI coding tools, agentic automation products, or reasoning-heavy internal tools now have another large, permissively licensed option to weigh against both closed APIs and other open checkpoints. If Reflection’s efficiency claims hold up under independent testing, Beam could pressure the pricing of comparable closed models by giving self-hosting enterprises a credible lower-cost alternative for reasoning and coding workloads specifically, the same workloads that tend to generate the heaviest API bills at scale.
There is also a capital markets angle. Reflection is a well-funded AI lab, and a release of this scale, positioned explicitly around competing with Chinese open models, lands at a moment when investors and policymakers are paying close attention to which country’s AI labs lead the open-weight ecosystem. Fortune’s framing of the release asked directly whether Beam could be “America’s best chance to beat China” in open models, a question that captures how much weight is being put on a single model launch as a proxy for a much larger competitive narrative.
What This Means for Enterprises and Developers
For teams currently building on top of open-weight models, Beam is worth tracking for three reasons. First, the Apache 2.0 license removes a common legal blocker to commercial adoption, so procurement and legal review should be straightforward once the weights ship. Second, the 1-million-token context window gives agentic and coding use cases more room to work with large codebases, long chat histories, or retrieved document sets without aggressive chunking. Third, and most consequentially if verified, the inference efficiency claim could change the cost calculus for self-hosting a frontier-scale reasoning model versus paying per-token for a closed API.
None of that is actionable yet in production, however, since Beam remains in a waitlist-gated early access phase as of this announcement. Teams evaluating it should plan to wait for the promised technical report and model card before committing engineering time, since those documents will carry the benchmark methodology needed to actually test Reflection’s efficiency claims rather than take them at face value.
Predictions: Where Beam and the Open-Weight Race Go Next
- Independent benchmark groups and rival labs will move quickly to test Reflection’s 3-4x inference-compute claim once weights are public, and the number will likely narrow once real-world serving conditions, rather than Reflection’s own test setup, are applied.
- Expect head-to-head cost comparisons between Beam and other large 2026 open releases, including Mistral Large 4 and Meta’s Muse Spark Open, to become a standard benchmark exercise for teams evaluating self-hosted reasoning models.
- The Apache 2.0 license will likely draw fast fine-tuning and distillation activity from startups looking to build smaller, specialized coding or agent models on top of Beam’s base weights.
- Reflection’s explicit “Western open ecosystem” framing suggests more US and European labs will lean into similar competitive messaging against Chinese open models in upcoming releases, turning efficiency claims into a recurring marketing battleground alongside raw parameter counts.
- If the technical report due later this month substantiates the efficiency numbers, expect enterprise interest in self-hosted agentic deployments to accelerate, since compute cost has been one of the biggest blockers to running large reasoning models continuously in production.
Frequently Asked Questions
What is Reflection’s Beam model?
Beam is an open-weight, sparse Mixture-of-Experts AI model released by Reflection with 501 billion total parameters and 23 billion active parameters per token, built for coding, reasoning, and agentic workloads.
How many parameters does Beam have?
Beam has 501 billion total parameters, but its sparse architecture activates only 23 billion of them for any given token, which is what keeps inference costs lower than a dense model of comparable total size.
Is Beam open source?
Yes. Reflection released Beam under an Apache 2.0 license, a permissive open-source license that allows commercial use, modification, and redistribution.
When will Beam’s weights be released?
At the time of the announcement, Beam was undergoing final red-teaming and evaluations, with only waitlist-based early access available. Reflection said weights, a technical report, a model card, and developer artifacts were planned for release later in the announcement month.
What is a Mixture-of-Experts model and why does Beam use one?
A Mixture-of-Experts (MoE) model splits its parameters into specialized sub-networks, called experts, and routes each token through only a subset of them. Beam uses 256 experts per MoE layer across 52 layers, routing each token through 8 of them, which lets the model carry a large total parameter count while keeping active compute, and therefore inference cost, much lower.
How does Beam compare to DeepSeek, Mistral, and other open models?
Beam’s 501 billion total parameters put it below Mistral Large 4’s 1.05 trillion parameters but above Meta’s Muse Spark Open at 816 billion and Nvidia’s Nemotron 3.5 Super VL at 120 billion. Reflection claims Beam uses 3-4x less inference compute than leading Western open models, including comparisons implied against models like DeepSeek’s releases, though that specific efficiency claim has not been independently verified.
Can I use Beam for commercial projects?
The Apache 2.0 license Reflection attached to Beam permits commercial use once the weights are publicly released, though early access during the waitlist phase may carry its own separate terms that have not been detailed publicly.
Is Reflection’s efficiency claim verified independently?
No. As of the announcement, the claim that Beam uses 3-4x less inference compute than leading Western open models was a statement from Reflection itself and had not been confirmed through independent, third-party benchmarking.
