Reflection’s Beam has interesting numbers, but you can’t download it yet, so for now it’s a model to watch rather than one to plan around. If you need an open-weight model today, Qwen 3.8 is the practical choice among the three: its weights are already published, and the 27B version is realistic to run yourself. I’d revisit Beam once the weights and licence are actually released and someone other than the vendor has tested it.
Live API prices: current per-token prices for DeepSeek, Qwen and more than 400 other models are in the AI models table on our dashboard, read twice a day from OpenRouter’s public model list. Use the provider filter to compare them.
I’ve tried DeepSeek myself, but not in a way that lets me compare it with these benchmark numbers, and I haven’t used Qwen or Beam. So this piece relies on published results and on what I could check on GitHub.
Updated 15 October 2026. This is a source-based analysis: I haven’t run Beam, DeepSeek or Qwen, and every benchmark number below is vendor-reported unless I say otherwise.
Which One Can You Actually Use?
| Reflection Beam | DeepSeek V4.1 Flash | Qwen 3.8 | |
|---|---|---|---|
| Weights downloadable | Not yet (planned, Apache 2.0) | No public GitHub repository found | Yes, since 12–14 August 2026 |
| Size | 501B total, 23B active (MoE) | Not checked here | 2.4T-A95B and 27B versions |
| Strongest point | Claimed compute efficiency | Top scores on agentic coding tests | Available and self-hostable now |
| My verdict | Watch | Benchmark leader, check access terms | Practical choice today (27B) |
Beam in Brief
Reflection AI announced Beam on 5 October 2026: a sparse mixture-of-experts model with 501 billion parameters, of which about 23 billion are active per token, aimed at coding, reasoning and agents. It launched as a waitlisted API beta, with weights, documentation and deployment materials promised later in October under Apache 2.0.
- Text only, no image or audio input.
- Reasoning dial: low, medium (default), high, xhigh and max; reasoning can’t be switched off.
- Context: Reflection claims 1 million tokens after midtraining, but reinforcement learning ran at a 256,000-token maximum and the beta API caps requests at 262,144 tokens with output up to 131,072. On LongBench v2 it scores 65.5, close to Qwen 3.8 Max (66.3) and GLM 5.2 (64.0).
- API: model ID
Beam-501B-A23B; OpenAI-compatible for Chat Completions and model listing only (no Responses, Embeddings, Files or Batch). As of 6 October DataCamp found no published per-token price. - Distribution: on 6 October it wasn’t on Ollama, OpenRouter or Hugging Face.
Update, 17 October 2026: Beam is now listed on OpenRouter as “Beam 1.0 501B A23B”, added on 16 October, at $1 per million input tokens and $3 per million output tokens, with a listed context of about 1 million tokens. That’s OpenRouter’s price, which can differ from Reflection’s own API, and the listed context is larger than the 262,144-token cap of the beta API described above, so check the limit for the route you use. We haven’t confirmed whether the weights have been published yet. Current prices are in our AI models table.
The Benchmarks
Vendor-reported. Reflection sourced rivals’ scores from Artificial Analysis and DataCurve; I take them from DataCamp’s review, which I read in full. No independent lab had reproduced them in the sources I saw.
| Benchmark | Beam | GLM 5.2 | GLM 5.3 | Kimi K3 | DeepSeek V4.1 Flash |
|---|---|---|---|---|---|
| Terminal-Bench v2.1 | 80.1 | 81.0 | 88.2 | 88.3 | 90.6 |
| DeepSWE v1.1 | 44.4 | 44.0 | 61.0 | 68.0 | 74.2 |
| SWE-bench Pro v2-Hard | 77.2 | not reported | 84.3 | 88.2 | not reported |
| SWE Atlas Codebase QnA | 34.6 | not reported | 61.0 | 68.0 | not reported |
Beam’s best coding result is 80.9 on SWE-bench Verified, ahead of the 77.6 Reflection lists for Inkling and 70.7 for Nemotron 3 Ultra, and it edges GLM 5.2 on SWE-bench Pro v1 (65.5 vs 62.1). Against the models you’d realistically choose between, though, it trails on every agentic coding test where a score exists, and the 29.8-point gap to DeepSeek V4.1 Flash on DeepSWE is large.
Reasoning and tool use are mixed: 36.2 on Humanity’s Last Exam without tools (GLM 5.2 40.5, Kimi K3 46.9); 37.0 on AutomationBench (ahead of GLM 5.2’s 26.2, behind Kimi K3’s 46.7 and DeepSeek’s 54.8). Search is its weakest area: 77.4 on BrowseComp against 91.2 for Kimi K3, which an agent living on the open web would notice.
How to Read the Efficiency Claim
Reflection says Beam matches GLM 5.2 on advanced reasoning with three to four times less inference compute. On its Humanity’s Last Exam chart, Beam reaches 35.3% at an estimated 0.82 PFLOP per attempt; Inkling needs about 4.1 PFLOP for roughly 26%, and Nemotron 3 Ultra about 1.9 PFLOP for around 21%.
The method favours Beam. Compute is estimated as twice the active parameters times average tokens generated, which ignores prompt processing, long-context attention and serving overhead, and naturally flatters sparse models. Turing Post adds that the efficiency charts leave out Kimi K3, GLM 5.3 and DeepSeek V4.1 Flash. What would settle it is measured cost per solved task, and nobody has published that yet.
How Beam Was Built, According to Reflection
From Reflection’s announcement as summarised by DataCamp; vendor-reported and unverified.
- Architecture: 52 layers with many small experts, only a few active per token.
- Training: 23.8 trillion pretraining tokens on 6,144 NVIDIA GB300 GPUs, then reinforcement learning on 10,500 GPUs with over 100 million rollouts across about 1 million coding, agentic and STEM tasks.
- Effort control: a length penalty rewarded correct answers while discouraging extra tokens, hence the per-request reasoning dial.
- Safety: a separate safety model trained from the same checkpoint and merged in, plus adversarial rounds. Reflection plans to open-source its safety evaluations; they weren’t published in the sources I read.
The most interesting claim is transfer: browsing scores reportedly improved even though the training mix had no browsing tasks. That’s exactly the kind of result worth waiting for outside labs to confirm.
What We Could Verify on GitHub
Method: on 15 October 2026 we looked for each model’s official GitHub repository and, where one existed, cloned it and read its README, licence file and recent history. This was done with AI assistance. We could not download or run any of the models, because our environment blocks Hugging Face, so we have no benchmark results of our own.
| Model family | Official GitHub repository | What it shows |
|---|---|---|
| Qwen 3.8 | Yes (QwenLM/Qwen3.8; QwenLM/Qwen3.5 now carries the same README) | Open weights for Qwen3.8-2.4T-A95B on 12 August 2026 and Qwen3.8-27B on 14 August. The README calls it the first openly released “Qwen-Max-class” model. Code is Apache 2.0; the weights’ licence ships with the model files. Last README update 17 August 2026. |
| DeepSeek V4.1 | None found under deepseek-ai/DeepSeek-V4, DeepSeek-V4.1, DeepSeek-V4-Flash or DeepSeek-V4.1-Flash | The newest repository we could open, DeepSeek-V3, was last updated 28 August 2025 and uses separate licences for code and weights. |
| Reflection Beam | None found under the names we checked | Consistent with weights being planned, not yet released. |
That difference matters more than any benchmark row. You can download Qwen 3.8 today, read its licence and test it on your own hardware. You can’t do that with Beam yet, and for DeepSeek’s newest model I couldn’t find a public repository to check. Note the 2.4-trillion-parameter Qwen is far beyond ordinary hardware; the 27B version is what most teams could actually run.
Before You Trust Any Headline Score
- Same benchmark version, same tool access, same number of attempts?
- Is test-time compute disclosed?
- Was the model run with repository search and execution tools, or prompted in a chat window? Those aren’t comparable.
- Are the weights actually downloadable, or is it a hosted API with a launch announcement?
How I’d Test Them Myself
- Run identical repository bug-fix tasks with fixed budgets.
- Measure end-to-end latency, GPU memory and cost per successful task.
- Review the code changes for security and maintainability.
- Repeat with multiple seeds and publish failures as well as successes.
FAQ
Can I download Reflection AI Beam?
Not as of our 15 October 2026 check. Reflection said weights would arrive later in October; check its official announcement rather than unofficial download links. You can, however, call it through OpenRouter’s API since 16 October.
Is Beam better than DeepSeek for coding?
Not on the published numbers. DeepSeek V4.1 Flash leads Beam on Terminal-Bench v2.1 (90.6 vs 80.1) and DeepSWE (74.2 vs 44.4).
Which open-weight model can I run today?
Of these three, Qwen 3.8; the 27B version is the realistic self-hosting option.
Related: MiniCPM5-2B for local agents · System 1 vs System 2 AI agents · Gemini 4 Argon explained · Tools for running open models locally
Sources
- Reflection AI: Introducing Beam
- DataCamp: Beam, Reflection AI’s 501B open-weight coding model (6 October 2026)
- Turing Post: Reflection’s Beam and the global frontier (excerpt only)
- Help Net Security: Beam trails top open models on coding tests
- GitHub repositories QwenLM/Qwen3.8, QwenLM/Qwen3.5 and deepseek-ai/DeepSeek-V3, checked 15 October 2026
