Reflection Beam: 501B Open Model from Ex-DeepMind Team

👁 113 reads

Reflection AI, a US startup founded by two former Google DeepMind researchers, announced Beam on October 5, 2026 - their first open-weight model. It’s a mixture-of-experts with 501 billion parameters, of which 23 billion are active per token, trained on 23.8 trillion tokens, with a 1 million token context window and Apache 2.0 license. Weights are scheduled for release later in October. We checked what exactly is known, where the numbers are, and what’s still just promises.

Who’s Behind Beam

Reflection AI was founded in 2024 in New York. The founders are Misha Laskin, who led reward modeling for the Gemini project at Google DeepMind, and Ioannis Antonoglou, co-creator of AlphaGo. According to PitchBook data cited by TechCrunch, the company has raised approximately $4.7 billion total, and was valued at $25 billion pre-money in its latest round. Investors include Nvidia, Sequoia and Lightspeed. Reflection also has compute agreements with SpaceX and Nebius worth over $7 billion combined - we covered the SpaceX deal in SpaceX to rent chips to Reflection AI.

The company has been explicit from the start about Beam’s purpose: to be the Western answer to open models from China - DeepSeek, Qwen, GLM and Kimi - which have dominated open-weight model rankings for over a year. A day later, Mistral adopted a similar narrative with the launch of Mistral Large 4.

Beam by the Numbers

Interesting detail from the data description: according to Reflection, their quality classifiers rejected about 95% of raw tokens from the internet. That’s a lot for a model this size and probably explains why the company emphasizes “efficiency per token” so strongly.

Results According to Reflection

Below are numbers from the table in the announcement. These are results provided by the company itself, without independent verification - the weights aren’t out yet, so no one external could reproduce them.

A dash means Reflection didn’t provide the competitor’s score. What this shows:

• Beam doesn’t beat the top Chinese models. In every test with full comparisons, Kimi K3, GLM-5.3 and Qwen 3.8-Max score higher. Reflection mainly compares itself to the older GLM-5.2, with which it’s neck and neck (e.g. GPQA 90.5 vs 91.2).
• The main argument is cost, not quality. The company claims it uses 3-4x less compute than GLM-5.2 on reasoning tests. That’s the manufacturer’s estimate, not measured deployment cost.
• 23B active parameters is small. For comparison, GLM-5.2 has about 744B parameters and 40B active, and Mistral Large 4 has 1.05T and 49B active. Fewer active parameters means cheaper and faster responses on the same hardware.

Can You Run Beam Locally

On a home computer - no. 501B parameters in 16 bits is about 1 TB of weights alone, and even after quantization to 4 bits you’re left with about 250 GB (our calculation: number of parameters times half a byte). That’s server-level with multiple H200 or B200 cards, not a gaming GPU. Our RTX 4090 has 24 GB. The fact that only 23B are active speeds up computation, but all experts still need to sit in memory. We covered how much memory smaller models need in how much VRAM you need.

Realistically, Beam will reach regular users through API providers: Reflection promises distribution through major cloud providers and open source libraries. When it appears on OpenRouter, we’ll run it through the same 13 tasks in Polish as all models in our AI Model Arena in Polish.

Why We Don’t Have Our Test Yet

• No public API. Early access is waitlist-only, and we don’t create accounts just to skip the queue.
• No weights. Announced for “later in October”.
• Not on OpenRouter. We checked the full model list on October 7 morning - neither Reflection nor Beam are there.

Honestly: nothing is known today about how Beam writes in Polish. Reflection’s announcement doesn’t mention languages other than English at all, and the model is focused on code and agents.

Beam vs Other October Open Releases

Who this matters for: companies that want an open model for code and agents, but for various reasons can’t or don’t want to use models from China. For regular chat users, nothing changes yet - Beam doesn’t have its own app. We tested Chinese models in Polish in Chinese AI models in Polish.

What We’ll Check When Weights Drop

1. Whether the license in the repository is truly clean Apache 2.0, without additional conditions.
2. How much it costs at API providers compared to GLM-5.3 and Kimi K3.
3. How it writes in Polish - 13 tasks from our Lab, same blind judge evaluation as over 30 other models.
4. Whether the claimed 3-4x less compute translates to fewer reasoning tokens in practice.

Sources

• Reflection: Introducing Beam (Oct 5, 2026) - parameters, data, results table, license
• TechCrunch, Oct 5, 2026 - 23.8T tokens, funding, valuation
• SiliconANGLE, Oct 5, 2026
• MarkTechPost, Oct 5, 2026
• OpenRouter model list (API /models), checked Oct 7, 2026

Checked: October 7, 2026.

// read moreNvidia in Talks to Acquire $25B Startup Reflection AINewsLiquid AI d1: Open Decision Model, AA Benchmark, and PerformanceNewsOlmo-core 3: Open Framework for Training Trillion-Parameter MoEAI Tools

Kinetyra