🌏 中文版
Model Overview
| Field | Value |
|---|---|
| Model ID | reflection/Beam-1.0-501B-A23B |
| Vendor | Reflection AI |
| Parameters | 501.3B total / 23.4B active (MoE, 256 experts, 52 layers: 3 dense + 49 MoE) |
| Context Window | 1,000,000 tokens |
| Input Price (USD/1M tokens) | $0.40 (hosted by Together AI; $0.04 cached) |
| Output Price (USD/1M tokens) | $1.20 (hosted by Together AI) |
| Open Weight | Yes (Apache 2.0) |
| Release Date | 2026-10-16 (weights, model card and safety report; teased 2026-10-05) |
| Official Announcement | Reflection AI Blog: Introducing Beam |
| HuggingFace | reflection/Beam-1.0-501B-A23B |
| Family | Beam (Reflection AI's first model family; the company says larger follow-ups are already in training) |
Capability Highlights
- A sparse MoE architecture (501.3B total / 23.4B active, 256 experts, gated GQA + sliding-window attention) with a 1M-token context, built specifically for agentic coding on text-only input and output
- Reflection claims comparable solution quality to peer open models while using roughly 3x fewer reasoning tokens, which translates into higher throughput per GPU-hour
- Pretrained on 23.8 trillion tokens, with an RL stage that ran 10,500 NVIDIA GB300 GPUs for four weeks and produced more than 100 million rollouts — reportedly one of the largest RL runs by any open lab to date
- Shipped in three numeric precisions (BF16, FP8, NVFP4) alongside a speculative-decoding drafter model, which Reflection says is unusual among open-weight releases
Benchmark Results
| Benchmark | Beam (BF16) | Prior Generation | Best Competitor (per Reflection's own preview chart) |
|---|---|---|---|
| SWE-bench Verified | 81.6% | No predecessor to compare | Inkling 77.6%, Nemotron 3 Ultra 70.7% (GLM/Kimi/Qwen were not reported on this bench) |
| Terminal-Bench v2.1 | 80.1% | No predecessor to compare | DeepSeek V4.1 Flash 90.6%, Kimi K3 88.3%, GLM 5.3 88.2% |
| SWE-bench Pro v1 | 66.1% | No predecessor to compare | Qwen 3.8 Max 67.7%, GLM 5.2 62.1% |
| AIME 2026 | 97.7% | No predecessor to compare | ⚠️ No competitor score was published on the same benchmark |
⚠️ All figures are Reflection's own self-reported evaluations. The "best competitor" column is pulled from the comparison chart in Reflection's October 5 early-access blog post; its evaluation settings differ slightly from the final October 16 model card, so treat the numbers as indicative rather than independently reproduced. The FP8 and NVFP4 checkpoints score 80.2% and 80.7% on SWE-bench Verified respectively — within 1.5 points of BF16.
Versus Predecessor / Competitors
Beam is Reflection AI's first open-weight model, so there is no predecessor to benchmark against — but the company put it straight into the ring with China's open-weight flagships. Reflection's own blog says Beam is "competitive with larger open models like GLM 5.2" and "approaching Qwen 3.8-Max," while conceding that models like Kimi K3 remain ahead on raw capability: on Terminal-Bench v2.1, both Kimi K3 (88.3%) and DeepSeek V4.1 Flash (90.6%) clearly beat Beam's 80.1%.
Beam's real pitch isn't the leaderboard — it's efficiency. With only 23.4B of 501.3B total parameters active per token (a 4.7% activation rate) and Reflection's claimed "3x reasoning-token efficiency," Beam reaches 81.6% on SWE-bench Verified at a fraction of the inference cost a 2T+-parameter model like Qwen 3.8 Max would require. Together AI's hosted pricing of $0.40 input / $1.20 output per million tokens sits in the low-to-mid range, but that reflects only the 23.4B active parameters — it says nothing about the cost of storing and deploying the full 501.3B weights, which still need roughly 251GB of memory even at 4-bit NVFP4.
What This Means for Agent Development
Beam's significance has less to do with topping benchmarks and more with who gets to use it: it's one of the few US-built models above 500B parameters released under Apache 2.0, and Reflection open-sourced its own coding-agent environment, Mirror CLI, alongside it — compatible with existing agentic harnesses like Claude Code, OpenCode, Pi, and Hermes, with native MCP tool calling and a 1M-token context.
- If you're building agentic coding tools and don't want to be locked into a single closed API: Beam's Apache 2.0 weights give you a self-hostable, fine-tunable, distillable path that doesn't force a choice between "Qwen/GLM/Kimi" and "Claude/GPT"
- If you're building agents that need long multi-step tool use — extended debugging sessions, repo-scale refactors: the combination of 1M context and token efficiency should, in theory, lower the inference cost of a single long-running task, and is worth a proof of concept
- Not a good fit for: workloads that are primarily non-English (Reflection states the model is aligned predominantly for English), ultra-low-latency single-turn chat (Beam's advantage is token efficiency over long tasks, not single-response speed), or teams without serious self-hosting infrastructure (even at 4-bit quantization, roughly 251GB of weights is a non-trivial deployment bar)
Today's Takeaway
Beam's official narrative splits the "open-weight frontier" into two separate races: Chinese labs are competing to push the raw benchmark ceiling higher (Kimi K3, GLM 5.3), while Reflection chose to compete on "same solution quality, fewer tokens." That's a reminder that judging open models purely by leaderboard position misses the variable that actually drives real-world inference cost — active parameters and reasoning-token count, not total parameter count or a single headline score.