Reflection Beam vs DeepSeek, Kimi and GLM: What the Benchmarks Show
By Matt Cullerton, founder of Mavric and Apso
Beam is Reflection AI's first open-weight language model, a 501-billion-parameter mixture-of-experts system announced on 2026-10-05. A TikTok from @thesanj23 asks whether it will "take down DeepSeek, Kimi, and GLM." Reflection's own benchmark table answers no: Beam trails GLM 5.3, Kimi K3 and DeepSeek V4.1 on nearly every coding and reasoning row, and the weights had not been released when the video said they had.
What is Reflection Beam?
Beam is a text-only sparse mixture-of-experts model with 501 billion total parameters and 23 billion active per token, a one-million-token context window, and 23.8 trillion tokens of pretraining. Reflection reports pretraining on 6,144 NVIDIA GB300 NVL72 GPUs in under four weeks, then a reinforcement-learning run on 10,500 GB300s for four weeks that produced more than 100 million rollouts at up to 256K context. The announced license is Apache 2.0.
Has Beam been released?
Not yet. The video says "today Reflection released Beam." Reflection's post says it "will release the weights, technical report, model card, and developer artifacts later this month," with the model "undergoing final red-teaming and evaluations" and an early-access waitlist in place. On announcement day, what shipped was a blog post, a benchmark table and a signup form.
Does Beam beat DeepSeek, Kimi and GLM?
The clip shows a cropped version of Reflection's table with only the Beam and Inkling columns visible. The full table on Reflection's blog includes the Chinese models, and it puts Beam behind them on most rows.
On agentic coding and terminal work:
- DeepSWE v1.1: Beam 44.4, against DeepSeek V4.1 at 74.2, Kimi K3 at 68.0 and GLM 5.3 at 61.0.
- Terminal Bench v2.1: Beam 80.1, below GLM 5.2 at 81.0, GLM 5.3 at 88.2, Kimi K3 at 88.3 and DeepSeek V4.1 at 90.6.
- SWE Bench Pro v2-Hard: Beam 77.2, behind GLM 5.3 at 84.3 and Kimi K3 at 88.2.
- SWE Bench Pro v1: Beam's 65.5 beats GLM 5.2's 62.1 but loses to Qwen 3.8 Max at 67.7. This is the only coding row where Beam leads a Chinese model outright.
On reasoning, Beam's 36.2 on HLE with no tools sits below GLM 5.2 (40.5), GLM 5.3 (42.3), Kimi K3 (46.9), Qwen 3.8 Max (43.6) and DeepSeek V4.1 (39.1). On GPQA Diamond, Beam's 90.5 is the lowest of the six reporting models except Nemotron 3 Ultra.
Reflection does not claim to lead. Its own wording is that Beam "achieves scores comparable to GLM-5.2 while using 3-4x less inference compute." That compares Beam to the previous GLM generation and makes an efficiency claim.
How reliable are Beam's benchmark numbers?
Every number is self-reported. Reflection ran the evaluations, chose the comparison set, and decided which cells to mark NR (not reported). No third party had reproduced the table as of 2026-10-06.
The efficiency claim carries a bigger caveat. Reflection's methodology note says the estimates "exclude prompt prefill, context-dependent attention operations, and serving overhead, so they represent an approximate compute comparison rather than measured inference cost." The 3x to 4x figure is arithmetic on active parameters and generated tokens. Prefill and attention are where long-context agentic workloads spend most of their compute, and both are excluded.
Is Beam smarter than Nvidia's Nemotron Ultra?
Mostly, on Reflection's figures. The video says Beam is "more intelligent than Nvidia's NemoTron Ultra model." The model is Nemotron 3 Ultra, a 550-billion-parameter open-weight model with roughly 55 billion active per token, announced at Computex on June 4, 2026 with weights, training data and recipes published.
Beam leads it on SWE Bench Pro v1 (65.5 vs 46.4), Terminal Bench v2.1 (80.1 vs 56.4), SWEBench Verified (80.9 vs 70.7), SWEBench Multilingual (78.0 vs 67.7), HLE (36.2 vs 26.7), GPQA Diamond (90.5 vs 87.0), MCP Atlas (78.7 vs 63.1) and BrowseComp (77.4 vs 44.4). Nemotron 3 Ultra leads on IFBench (81.7 vs 79.7), and the two tie on AA-LCR at 79.3. The result is strong for Beam on coding and agentic work but not a sweep.
Is Beam the first US open-weight option?
The strongest point in the video is about procurement: "a lot of my customers cannot get approvals internally to be able to fine tune and serve models that are created in China." That constraint is real.
Beam is not the first US option for those buyers, though. The "Inkling" column visible in the video's own screenshot is Thinking Machines Lab's 975-billion-parameter open-weights model with 41 billion active, from a US lab founded by Mira Murati. Nemotron 3 Ultra, which the video names, is a US open-weight release from June 2026. Beam adds a third choice.
The Chinese models do predate Beam. GLM 5.3 shipped on 2026-08-14 from Z.ai, formerly Zhipu AI. Kimi K3 shipped on 2026-07-16 from Moonshot AI. DeepSeek V4.1 shipped in September 2026. All three lead Beam on most published rows.
Reflection's table also contains only open-weight models, so it says nothing about how Beam compares with closed models from Anthropic, OpenAI or Google.
Is this the same Reflection as Reflection 70B?
No. Reflection AI was founded in 2024 by Misha Laskin and Ioannis Antonoglou, both former Google DeepMind researchers. It emerged from stealth in March 2025 with $130 million, raised $2 billion at an $8 billion valuation in October 2025, and closed a $2.5 billion Series C at a $25 billion pre-money valuation in April 2026.
Reflection 70B was a 2024 Llama fine-tune announced by Matt Shumer of OthersideAI. Independent evaluators could not reproduce its claimed scores, and Artificial Analysis measured the released weights as roughly on par with Llama 3. The two are unrelated. That episode is a good reason to treat any launch-day benchmark table as provisional until outside testers run the weights.
Bottom line
Beam is a real and large US open-weight model, and its scores match what Reflection published. By Reflection's own table it trails GLM 5.3, Kimi K3 and DeepSeek V4.1 on most coding and reasoning rows, and its efficiency figure excludes prefill and serving costs. Wait for the Apache 2.0 weights and independent results before choosing it on benchmarks.
Sources
- Introducing Beam: Reflection's announcement with the full benchmark table and methodology note.
- Reflection on X: Introducing Beam: launch post stating weights will release later in the month.
- CNBC interview with Misha Laskin: the CEO on the Beam launch, 2026-10-06.
- Inkling: Our Open-Weights Model: Thinking Machines Lab's announcement.
- Nemotron 3 Ultra coverage: model size and Computex announcement date.
- Reflection AI on Wikipedia: founders and funding history.
- VentureBeat on Reflection 70B: the unrelated 2024 episode.
Published October 5, 2026. Based on a TikTok by @thesanj23. Full video notes (markdown)