Reflection released Beam on October 16, 2026, its first open-weight language model, and handed the weights to the public under an Apache 2.0 license. The announcement, confirmed by Reflection’s own blog post alongside coverage from Unite.AI and DataCamp, marks one of the largest open-weight releases of the year by raw parameter count. Beam uses a sparse Mixture-of-Experts design with 501 billion total parameters, though only 23 billion are active during any given inference pass, a ratio that lets the model aim for frontier-level reasoning without frontier-level compute bills.
For a company best known until now for closed, API-gated research, the shift to Apache 2.0 licensing is notable on its own. Apache 2.0 is about as permissive as open licensing gets: it allows commercial use, modification, and redistribution with minimal restriction, which is exactly why enterprises favor it over more research-only licenses. Reflection paired the weight release with the tokenizer, a reference implementation, and the training harness itself, a combination that goes further than most “open-weight” launches, which frequently withhold the tooling needed to actually reproduce or fine-tune the model.
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 Reflection Actually Shipped on October 16
The Beam release bundles four distinct pieces: the model weights, the tokenizer, a reference inference implementation, and a companion component called DFlash2, described in the release materials as a drafter model used to accelerate generation. Reflection also open-sourced Mirror CLI, the command-line harness it says it used internally to build and train Beam. Releasing the training harness alongside the weights is unusual; most labs keep their training stack proprietary even when they open the weights, because the harness often reveals more about how a model was actually built than the weights themselves.
Reflection says Beam targets coding, reasoning, and agentic workloads specifically, rather than positioning the model as a general chatbot replacement. That framing lines up with where most of the commercial demand for large open-weight models currently sits: engineering teams that want a model they can host, fine-tune, and audit internally, instead of routing every request through a third-party API. The company’s own description of the launch, carried in its blog post, states plainly that “Beam is now available across platforms, with open weights under the Apache 2.0 license.”
Inside Beam’s Sparse Mixture-of-Experts Architecture
Beam’s 501-billion-parameter total count puts it in the same size class as some of the largest open releases this year, including Mistral Large 4, which launched at 1.05 trillion parameters. But raw parameter count tells only part of the story with a Mixture-of-Experts model. Because Beam only activates 23 billion parameters per token, the effective compute cost of running the model looks much closer to a mid-sized dense model than to a 501-billion-parameter monolith. That is the entire point of the sparse MoE approach: route each token to a small subset of specialized expert sub-networks rather than running every parameter on every request.
This isn’t a new idea. Mixtral, DeepSeek’s V-series, and Nvidia’s Nemotron line have all used variations of sparse MoE routing to stretch parameter budgets further than dense architectures allow. What makes Beam notable is less the architecture choice itself and more that Reflection chose to open it fully, including the routing and training code, at a scale few open releases have reached. Nvidia’s own open Nemotron family, by comparison, tops out well below Beam’s total parameter count in its largest variant.
Apache 2.0: Why the License Matters as Much as the Model
Licensing terms shape adoption as much as benchmark scores do, and Apache 2.0 is the license enterprise legal teams tend to clear fastest. Unlike more restrictive open-weight licenses that carry usage caps tied to company revenue or user counts, Apache 2.0 imposes no such ceiling. A startup and a Fortune 500 company face the same terms. That matters for a model aimed squarely at coding and agentic workloads, where companies often want to fine-tune on proprietary codebases and deploy the result behind their own firewall without negotiating a separate commercial agreement.
Reflection’s Dr Hrant Kostanyan framed the release in similarly direct terms, stating: “Today we released Beam, Reflection’s first open-weight model, under Apache 2.0.” The brevity of that statement is itself a signal. Where some labs wrap open releases in lengthy safety disclaimers or staged rollout plans, Reflection’s framing treats the Apache 2.0 license as the headline, not a footnote.
DFlash2: The Drafter Model Behind Beam’s Speed Claims
Alongside Beam itself, Reflection released DFlash2, described in the announcement as a drafter model. Drafter models are typically paired with a larger model in a technique known as speculative decoding, where the small drafter proposes several tokens ahead and the larger model verifies them in a single pass, cutting the number of expensive full-model forward passes needed to generate a response. Pairing a dedicated drafter with a 501-billion-parameter MoE model suggests Reflection is optimizing specifically for inference latency, a priority that tracks with Beam’s stated focus on agentic workloads, where a model might need to make dozens of tool calls in sequence and any added latency compounds quickly.
Reflection has not published independent benchmark figures tying DFlash2 to a specific speedup multiplier in the materials reviewed for this story, so any specific throughput claim beyond what the company has stated should be treated as unconfirmed until third-party benchmarking firms publish their own numbers.
Mirror CLI: Opening the Training Harness, Not Just the Weights
Reflection describes Mirror CLI as the open-source harness it used to build and train Beam. Open-sourcing the actual training tooling, rather than just the resulting weights, is a step beyond what most labs do when they announce an “open” model. Meta’s Llama releases, for instance, have historically shared weights and some technical reporting without releasing the full internal training stack. Reflection’s choice to ship Mirror CLI gives outside developers and researchers a closer look at how Beam was actually produced, which matters for reproducibility and for any academic or independent audit of the model’s training process.
Reflection says the release involved an ecosystem of more than 40 partners, though the company’s materials do not break down exactly what role each partner played, whether that is infrastructure, distribution, fine-tuning support, or something else. The scale of that partner list, regardless of the specifics, signals that Reflection built day-one distribution into the launch rather than treating it as a solo release.
Weight Formats: BF16, FP8, and NVFP4 Explained
Reflection shipped Beam in three separate weight formats: BF16, FP8, and NVFP4. Offering multiple precision formats at launch is a practical acknowledgment that not every team running Beam has the same hardware or latency budget. BF16 is the higher-precision, higher-memory option typically used when accuracy matters most and hardware constraints are looser. FP8 trades some precision for roughly half the memory footprint and faster throughput on hardware that supports it. NVFP4 is a newer, even more compressed format built for Nvidia’s latest accelerators, aimed at squeezing inference cost down further at the expense of some numerical precision.
Shipping all three at launch, rather than BF16 alone with quantized versions to follow later, suggests Reflection anticipated demand from teams running on a range of hardware, from older BF16-optimized clusters to newer Nvidia silicon built around lower-precision formats.
| Weight Format | Typical Use Case | Relative Memory Footprint | Precision Trade-off |
|---|---|---|---|
| BF16 | Maximum accuracy, research and evaluation | Highest | Lowest numerical loss |
| FP8 | Production inference on supporting hardware | Roughly half of BF16 | Moderate |
| NVFP4 | Cost-optimized inference on newer Nvidia accelerators | Lowest | Highest compression, most aggressive |
| Reference implementation | Reproducibility and custom deployment | N/A | N/A |
| DFlash2 drafter | Speculative decoding to speed up generation | Small relative to Beam itself | N/A |
How Beam Stacks Up Against the Rest of 2026’s Open-Weight Field
Beam arrives into a crowded open-weight landscape that has moved fast throughout 2026. Meta’s Muse Spark Open launched on seven clouds simultaneously, while Switzerland’s Apertus 1.5 pitched itself as a fully open alternative with a long context window. Microsoft’s MAI-Base-2 went even larger on paper, at 2.8 trillion parameters, while smaller players like Liquid AI’s LFM2.5 chased efficiency at a fraction of the size. Beam’s 501 billion total parameters with 23 billion active places it in an odd middle ground: large enough to compete with flagship closed models on paper, sparse enough to run more cheaply than its total parameter count would suggest.
The closed-model side of the market has not stood still either. Google’s Gemini 4 Argon went public with its own pricing structure just weeks before Beam’s release, and Anthropic’s Claude Opus 5.5 remains a reference point for reasoning-heavy workloads. Every new open-weight release at this scale puts pressure on the closed labs to justify their pricing, since a capable, permissively-licensed alternative gives enterprise buyers real leverage in that negotiation, even if actual benchmark parity has not been independently verified.
| Model | Developer | Total / Active Params | License | Release Window |
|---|---|---|---|---|
| Beam | Reflection | 501B / 23B | Apache 2.0 | October 2026 |
| Mistral Large 4 | Mistral AI | 1.05T total | Commercial/open mix | 2026 |
| MAI-Base-2 | Microsoft | 2.8T total | Not Apache 2.0 | 2026 |
| Nemotron 3.5 family | Nvidia | Up to 120B (Super VL variant) | Open | 2026 |
| LFM2.5 | Liquid AI | 24B total | Open | 2026 |
| Apertus 1.5 | Swiss AI initiative | Not disclosed in this comparison | Fully open | 2026 |
Note that context length for Beam is unconfirmed in the official sources reviewed for this story, so it is deliberately left out of the comparison above rather than estimated.
Pricing and Access: What’s Confirmed and What Isn’t
Reflection’s own release materials do not list a universal price for accessing Beam, which is typical for an open-weight release: the whole point is that anyone with sufficient hardware can self-host the model rather than paying a per-token API fee. That said, third-party hosting platforms have begun listing their own pricing for teams that would rather rent inference than run it themselves. One such listing, reported for the Baseten Model API, puts Beam access at $1.30 per million input tokens, $4.30 per million cached-input tokens, and $0.13 per million output tokens. That figure comes from a third-party hosting provider rather than Reflection directly, so it should be read as one data point on API-style access rather than a confirmed, universal price for Beam.
The split between a free, self-hosted path and a paid, managed-hosting path is the same model the rest of the open-weight market has settled into. Weights are free to download and run under Apache 2.0; convenience, in the form of managed inference, scaling, and support, is where providers like Baseten make their money. AlphaXiv’s listing for the release reflects that same pattern, tracking the open paper and weights separately from any commercial hosting layer.
Why a 40-Plus Partner Ecosystem Matters at Launch
Reflection’s announcement cites an ecosystem of more than 40 partners tied to the Beam launch. For context, most open-weight releases arrive with far less coordinated day-one support, typically a handful of inference providers adding the model to their catalog in the days or weeks following release. A 40-plus partner count at launch suggests Reflection spent meaningful time lining up distribution before going public with the weights, rather than releasing first and waiting for the ecosystem to catch up organically.
That kind of coordinated rollout matters commercially because the biggest barrier to adoption for any new open-weight model isn’t usually interest, it’s availability on the platforms where developers already work. A model that shows up simultaneously across many hosting providers, fine-tuning platforms, and inference engines gets evaluated by far more teams in its first week than one that trickles out through a single official channel.
Market Impact: Pressure on Closed-Model Pricing
Every large, permissively-licensed open-weight release shifts the negotiating position of enterprise AI buyers, even when they never actually switch away from a closed model. The mere existence of a credible, self-hostable alternative at Beam’s scale gives procurement teams a reference point when negotiating enterprise API contracts with closed-model vendors. That dynamic has already played out around other 2026 releases: OpenAI’s retirement of GPT-5.5 in favor of newer models came amid exactly this kind of competitive pressure from both closed rivals and the open-weight field.
For infrastructure providers, a new large MoE model with multiple quantization formats also creates fresh demand for the hardware best suited to run it efficiently, particularly accelerators that support FP8 and NVFP4 natively. That’s a dynamic Nvidia in particular has leaned into with its own open releases, and Beam’s format choices suggest Reflection built the release with that same hardware landscape in mind rather than targeting a narrower set of accelerators.
Historical Context: From Llama to Beam
Open-weight releases have escalated in scale almost every year since Meta’s original Llama models pushed the idea into the mainstream. Each subsequent wave, from Mistral’s early releases through DeepSeek’s V-series and now the current crop including Mistral Large 4, Microsoft’s MAI-Base-2, and Reflection’s Beam, has pushed parameter counts higher while gradually loosening license terms toward the Apache 2.0 standard Beam now carries. What’s shifted more recently is which companies are participating. Reflection built its early reputation on closed research rather than open releases, so Beam represents a strategic pivot as much as a technical one, placing the company in direct competition with labs that have built their brand specifically around open weights.
The timing also matters. Beam landed in the same stretch of 2026 that saw widespread coverage of several other major open and closed releases in quick succession, a pace that has become the norm rather than the exception this year. Enterprise AI teams now effectively have to re-evaluate their model stack on a near-monthly basis, a cadence that was unthinkable just two or three years ago.
Developer Access: Getting Started With Beam
For teams wanting to experiment with Beam directly rather than through a hosted API, the reference implementation Reflection released alongside the weights is the starting point. A typical self-hosted workflow for a model released this way looks something like the command below, though exact flags will depend on which weight format and hardware target a team chooses.
# Example self-hosted workflow pattern for an Apache 2.0 MoE release
mirror-cli download beam --format fp8
mirror-cli serve beam --weights ./beam-fp8 --drafter dflash2
mirror-cli benchmark beam --task coding --task agenticTeams without the hardware to self-host at this scale will more likely reach Beam through a managed provider instead, where the model shows up as just another entry in an existing model catalog, billed per token rather than per GPU-hour.
What’s Still Unconfirmed About Beam
Two important specs remain unconfirmed in the official release materials reviewed for this story: Beam’s context window length and a universal, Reflection-set price for API access. Both are the kinds of details that typically surface in the days following a major release, as independent benchmarking outlets and hosting providers publish their own testing. Readers evaluating Beam for a specific workload should treat any context-length or pricing figure circulating outside Reflection’s own materials as provisional until the company confirms it directly or a wider set of named benchmarking outlets converges on the same number.
Predictions: Where the Open-Weight Race Goes Next
- Expect at least one major closed-model lab to respond to Beam with either a price cut or an accelerated release timeline within the next one to two months, continuing the pattern already seen around other 2026 open-weight launches.
- Hosting providers beyond Baseten will likely publish their own Beam pricing within weeks, and those figures may diverge meaningfully from the $1.30/$4.30/$0.13 per-million-token structure currently reported for Baseten’s Model API.
- Reflection’s decision to open Mirror CLI alongside the weights will likely push other labs to reconsider how much of their training tooling they keep closed, since reproducibility has become a bigger selling point for enterprise buyers evaluating open models.
- Given Beam’s stated focus on coding and agentic workloads, expect fine-tuned community variants targeting specific developer tools and IDE integrations to appear within the first one to two months post-release.
- The still-unconfirmed context length will likely become a point of comparison once independent testing publishes figures, and if it lands short of rivals like Apertus 1.5, that gap will shape which workloads teams route to Beam versus competing open models.
Why This Release Is Different From a Typical Model Drop
Most model announcements lead with benchmark charts. Reflection’s Beam announcement leads with licensing and tooling instead, the Apache 2.0 terms, the open training harness, and the multi-format weight release. That ordering is itself informative about who Reflection is trying to reach first: not casual users comparing chatbots, but engineering teams deciding whether to build production infrastructure around a model they can fully own, audit, and modify without a vendor relationship sitting in the middle. Whether Beam’s actual task performance holds up against that framing is something only independent benchmarking in the coming weeks will settle.
Frequently Asked Questions
What is Reflection’s Beam?
Beam is Reflection’s first open-weight language model, released October 16, 2026, as a sparse Mixture-of-Experts model with 501 billion total parameters and 23 billion active parameters per token, aimed at coding, reasoning, and agentic workloads.
What license does Beam use?
Beam is released under the Apache 2.0 license, one of the most permissive open-source licenses available, allowing commercial use, modification, and redistribution without revenue-based usage caps.
What weight formats are available for Beam?
Reflection released Beam in three formats: BF16, FP8, and NVFP4, covering a range of hardware and precision-versus-cost tradeoffs.
What is DFlash2?
DFlash2 is a drafter model Reflection released alongside Beam, intended to speed up generation through speculative decoding, where a smaller model proposes tokens that the larger model verifies in batches.
What is Mirror CLI?
Mirror CLI is the open-source training harness Reflection says it used to build and train Beam. Releasing it alongside the weights goes further than most open-weight launches, which typically withhold training tooling.
How much does it cost to use Beam?
Reflection has not published a universal price for API-style access. A third-party listing for the Baseten Model API reports $1.30 per million input tokens, $4.30 per million cached-input tokens, and $0.13 per million output tokens, but this figure is not confirmed as a universal Beam price. Self-hosting under Apache 2.0 carries no licensing fee.
What is Beam’s context length?
Context length is unconfirmed in the official release materials reviewed for this story. Readers should treat any specific figure circulating elsewhere as provisional until Reflection confirms it directly.
How does Beam compare to other open-weight models released in 2026?
Beam’s 501 billion total parameters sit between smaller efficiency-focused releases like Liquid AI’s LFM2.5 and larger releases like Mistral Large 4 and Microsoft’s MAI-Base-2. Its sparse MoE design, activating only 23 billion parameters per token, is intended to keep inference costs closer to a mid-sized dense model despite the large total parameter count.
