Reflection AI dropped Beam’s model weights on Hugging Face on October 16 under Apache 2.0. It’s a 501-billion-parameter sparse Mixture-of-Experts model with 23 billion active parameters per token, built for coding and long-horizon agentic work. You can download it today, self-host it, put it in a product, and nobody owes anyone a licensing call. That’s the headline.
The secondary headline, which most coverage is quietly burying: Beam benchmarks at roughly GLM 5.2 level on agentic coding tasks — a model from June 2026. China’s leading open models are still ahead by a significant margin. Knowing both things is how you decide whether to ship with it.
How to Start Using Beam Today #
The fastest path is the free API beta at platform.reflection.ai. It’s OpenAI-compatible, which means swapping it into an existing stack takes minutes:
from openai import OpenAI
client = OpenAI(
api_key="your-reflection-api-key",
base_url="https://api.reflection.ai/openai/v1"
)
response = client.chat.completions.create(
model="Beam-501B-A23B",
messages=[{"role": "user", "content": "Review and fix this Python function..."}],
extra_body={"reasoning_effort": "high"}
)
print(response.choices[0].message.content)
For self-hosting, the weights live on Hugging Face under Apache 2.0 in BF16, FP8, and NVFP4 formats. Day-zero inference support covers vLLM, SGLang, Hugging Face Transformers, and Unsloth. Day-zero support on AWS through Bedrock means you can deploy without leaving your existing cloud setup. The minimum self-hosting footprint is an 8-GPU server running 8-bit quantization — not trivial, but accessible to any team with a mid-tier GPU cluster.
The Benchmarks, Honestly #
Reflection’s press materials lead with SWE-bench Verified at 80.9 — a legitimate first place on that specific benchmark. When you look at the agentic coding tasks Beam was explicitly built for, the picture shifts:
| Benchmark | Beam | Kimi K3 | DeepSeek V4.1 Flash |
|---|---|---|---|
| SWE-bench Verified | 80.9 | — | — |
| Terminal-Bench v2.1 | 80.1 | 88.3 | 90.6 |
| DeepSWE v1.1 | 44.4 | 68.0 | 74.2 |
That 30-point gap on DeepSWE is real. It’s not a benchmark quirk. Reflection acknowledges Kimi K3 is ahead — which earns them credibility points — but the gap matters if your use case involves long-horizon agentic software development rather than isolated code fixes.
Reflection’s counter-argument is efficiency: Beam uses 3–4x less inference compute than GLM 5.2 to reach comparable scores on reasoning benchmarks. Co-founder Ioannis Antonoglou: “You don’t give up anything. You just penalize it in terms of how many tokens it can use.” That’s true within the comparison they chose. It excludes prompt processing costs and serving overhead, which shift the math for real deployments. Fewer tokens per attempt only beats more tokens if the fewer-token model doesn’t retry more.
The Self-Distillation Loop #
The most interesting part of Beam’s release isn’t the weights — it’s the post-training recipes. Reflection ships a documented self-distillation workflow:
- Generate Beam traces for your domain-specific tasks
- Verify successful traces using tests or reward signals
- Train a LoRA adapter on the verified examples
- Repeat — each iteration improves success rates and reduces token usage
For teams building vertical coding agents — internal tooling, domain-specific refactoring, regulated-industry development — this loop is what turns a general open model into a purpose-built one. The fine-tuning stack integrates with NVIDIA NeMo RL, Hugging Face TRL, Anyscale SkyRL, and Axolotl.
Mirror CLI and DFlash2 #
Two developer tools ship alongside the weights. Mirror CLI is an open-source terminal harness for interactive agent workflows — resumable conversations and switchable model endpoints. DFlash2 is a speculative-decoding drafter optimized for agentic workloads where models make many small sequential calls. In vLLM, DFlash2 cuts decode latency 2.71x. Both are worth evaluating before you build custom wrappers.
Beam also includes a reasoning effort dial — low, medium, high, xhigh, max — settable per request. Most coding tasks run well at medium or high; max is for the long-horizon problems where Beam will burn more tokens to think through multi-file refactors or complex debugging chains.
Who Should Use It Now #
Use Beam now if: you need Apache 2.0 for commercial deployment, you’re building for air-gapped or regulated environments where running a Chinese-origin model is off the table, or you want to run the self-distillation loop on a domain where GLM 5.2 already works well for you.
Wait if: your primary use case is long-horizon autonomous coding agents and you’re comparing against the best available models. The 30-point DeepSWE gap is significant enough that Kimi K3 or DeepSeek V4.1 Flash will produce meaningfully better results today. Beam’s second version will be the one to watch — Reflection has the compute budget and the team to close that gap, and they’ve now proven they can ship open weights.