Reflection's Beam is out as open weights: 501B parameters, Apache 2.0, and what it costs to run
Reflection AI released Beam's weights on 16 October. The specs, the two sets of benchmark numbers, self-hosting memory, and API prices by provider.
Reflection AI has published the weights of Beam, its first open-weight model, eleven days after announcing it. The launch post dated 16 October says the weights are on Hugging Face under Apache 2.0, with BF16, FP8 and NVFP4 checkpoints and a small drafter model for speculative decoding. That puts a 501B-parameter model under a standard permissive licence. It is also too big for most teams to run themselves, so the hosted prices matter more than the download.
This post covers what was released, the benchmark numbers (which changed between the announcement and the model card), what self-hosting involves, and what the API options cost today.
What was released
From Reflection’s model card:
- Size: sparse mixture-of-experts, 501.3B total parameters, 23.4B active per token. 52 layers (3 dense, 49 MoE) with 256 experts.
- Context: 1M tokens. Text in, text out; no image input.
- Knowledge cutoff: 30 June 2026.
- Languages: the card says Beam was developed “predominantly for English” and lists no other languages.
- Precisions: BF16, FP8 and NVFP4.
The earlier announcement post adds the training details: 23.8T pretraining tokens, a reinforcement-learning run of four weeks on 10.5K NVIDIA GB300 GPUs with more than 100 million rollouts, and an RL context of up to 256K tokens that midtraining extends to 1M.
Reflection’s API is in public beta and the launch post describes it only as “free to start”, with no per-token price. The same post lists more than 40 launch partners, including AWS, Google Cloud, Microsoft Foundry, Cloudflare, OpenRouter, Together AI, Fireworks, Ollama and LM Studio, with Dell and HPE for on-premises and air-gapped setups. The launch post promises self-distillation recipes “in the coming days”; they were not part of the 16 October release.
The benchmarks: two sets of numbers
Reflection published coding scores twice: in the announcement (dated 5 October, with results updated 8 October) and in the model card that came with the weights. They don’t match exactly, and the card doesn’t explain why. Both are Reflection’s own evaluations; we found no independent results yet.
| Benchmark | Announcement (8 Oct) | Model card, BF16 | Model card, NVFP4 |
|---|---|---|---|
| DeepSWE v1.1 | 44.4 | 43.3 | 42.7 |
| SWE-Bench Pro v2-Hard | 77.2 | 76.5 | 78.6 |
| SWE-Bench Pro v1 | 65.5 | 66.1 | 65.6 |
| Terminal-Bench v2.1 | 80.1 | 80.1 | 78.4 |
| SWE Atlas Codebase QnA | 34.6 | 31.8 | 31.4 |
| SWE-Bench Multilingual | 78.0 | 78.3 | 77.5 |
| SWE-Bench Verified | 80.9 | 81.6 | 80.7 |
Between the announcement and the BF16 column the differences are small, from 0 to 2.8 points. If you quote a number, say which source it came from. The more useful finding is in the last column: the 4-bit NVFP4 checkpoint loses at most 1.7 points against BF16 on these tests, and scores higher on one of them. For anyone self-hosting, that is the checkpoint to try first.
Where it stands against other open models
Reflection’s announcement compares Beam with other open models and doesn’t hide where it trails. On Terminal-Bench v2.1, Beam’s 80.1 is behind GLM 5.3 (88.2), Kimi K3 (88.3), Qwen 3.8 Max (86.6) and DeepSeek V4.1 Flash (90.6). On DeepSWE v1.1 the gap to DeepSeek V4.1 Flash is wide: 44.4 against 74.2. That table predates DeepSeek’s 17 October Flash update, so the Flash figures are for the earlier checkpoint.
Reflection’s argument is efficiency, not top scores. It says Beam reaches “scores comparable to GLM-5.2 while using 3–4× less inference compute” on advanced reasoning benchmarks. The page also says how that was estimated: FLOPs ≈ 2 × active parameters × mean generated tokens, leaving out prompt prefill, attention cost and serving overhead. Treat it as a rough guide to token usage, not a measured cost.
Running it yourself
Reflection doesn’t publish minimum hardware. The launch post says its own production serving runs on SGLang on NVIDIA Blackwell, and lists vLLM, SGLang and Hugging Face Transformers among the supported engines.
You can estimate the memory from the parameter count. At 2 bytes per parameter, 501B parameters in BF16 is about 1 TB of weights. FP8, at 1 byte each, is about 500 GB, and a 4-bit format is roughly 250 GB plus scaling factors. That covers the weights only. The KV cache for long contexts comes on top, and a 1M-token context is expensive. Even at NVFP4 this is a multi-GPU server job. The 23B active parameters make each token cheap to compute, but the whole model still has to be in memory.
For most Indian teams, self-hosting makes sense only if the data truly can’t leave your own servers. If that’s you, the Apache 2.0 licence and the Dell/HPE air-gapped options are the reasons to look at Beam.
Using it through an API
OpenRouter’s Beam page lists the model as released on 16 October with a 1M context, Apache 2.0, and four providers when we checked:
| Provider | Input (US$/M tokens) | Output (US$/M tokens) |
|---|---|---|
| Nebius Token Factory | 0.25 | 1.00 |
| Together | 0.40 | 1.20 |
| CoreWeave | 1.00 | 3.00 |
| Baseten | 1.30 | 4.30 |
The most expensive provider charges more than four times the cheapest, for the same weights. If you route through OpenRouter, check which provider you’re actually getting, and read that provider’s data-retention terms. Cache-read prices also differ by provider.
Fireworks offers Beam only as an on-demand dedicated deployment, not serverless, and its page marks function calling and fine-tuning as unsupported there. If your agent depends on tool calls, check the provider you pick supports them for this model.
What it means for teams in India
- English first. The model card claims nothing beyond English. If your product serves users in Hindi, Tamil or other Indian languages, test those cases yourself before relying on it. For code and agent work in English it fits the brief.
- The licence is the selling point. Apache 2.0 lets you fine-tune, quantise, self-host and ship commercially, with no revenue threshold or custom terms to check with your lawyers.
- Check it on your own agent tasks. Reflection tested Beam with its own Mirror CLI and with Pi, OpenCode and Hermes. Results in other harnesses may differ, so run your own evals before switching from a model you already trust.
- Watch the price spread. At US$0.25/US$1.00 per million tokens on the cheapest listed provider, Beam is cheap to try. Pin the provider once you’ve picked one, so your costs don’t jump when the router switches.
References
- Beam is now available — Reflection AI, accessed 2026-10-19
- Introducing Beam — Reflection AI, accessed 2026-10-19
- Beam model card — Reflection AI, accessed 2026-10-19
- Beam 1.0 501B A23B: providers and pricing — OpenRouter, accessed 2026-10-19
- Beam 1.0 501B A23B model page — Fireworks AI, accessed 2026-10-19
This week in AI, every Sunday
One email with the releases that mattered and what they mean in practice. No spam.