model release

Reflection unveils 501B-parameter Beam, Mistral previews 1T-parameter Large 4, both open-weight

TL;DR

Reflection introduced Beam, a 501B-parameter mixture-of-experts model with 23B active parameters. Mistral said it is finishing Mistral Large 4, a 1T-parameter multimodal model with 49B active parameters. Both companies plan open-weight releases in October, and both are positioning the models against Chinese open-weight leaders.

3 min read
0

Reflection introduced Beam, a 501-billion-parameter mixture-of-experts (MoE) model, on Monday. A day later, Mistral said it is finishing Mistral Large 4, a 1-trillion-parameter multimodal model. Both are open-weight releases aimed at narrowing the gap with Chinese models from Alibaba, Z.ai, Moonshot and DeepSeek, according to Axios.

Reflection Beam

  • Architecture: MoE with 501B total parameters and 23B active per token.
  • Weights: Reflection plans to release them later this month, along with tools for running, evaluating and fine-tuning the model.
  • Performance claims: Reflection says Beam is competitive with Z.ai's GLM-5.2 and approaches Alibaba's Qwen 3.8-Max on some coding and agentic tasks. Its own published benchmarks are mixed: Beam beats or roughly matches the Chinese models on some tests and trails them on others.
  • Efficiency claim: Reflection says Beam reaches reasoning performance comparable to GLM-5.2 with three to four times less compute. This is a company claim and has not been independently verified.

Mistral Large 4 ("Le Chonk")

  • Architecture: 1T total parameters, 49B active, multimodal.
  • Training: 4,000 Nvidia Grace Blackwell GPUs over two months, in Mistral's own European data centers.
  • Access: A moderated API is the initial route. A version with fewer restrictions and broader cybersecurity capabilities is going to select partners for testing.
  • Weights: Mistral plans to release them on Oct. 27, after further reinforcement learning and safety testing.
  • Positioning: Mistral VP of science Pierre Stock says the company believes Large 4 is the world's best open-weight model and can outperform closed models on certain tasks. He also conceded: "We're not there yet on the frontier."

Context window, pricing, benchmark scores and training cutoff dates were not disclosed for either model. Mistral Large 4 is still being finalized, and neither model's weights are public yet.

The safety trade-off

Public weights make safeguards easier to remove, a growing concern as models improve at cybersecurity tasks. Both companies argue the opposite case. Reflection CEO Misha Laskin says a broad ecosystem inspecting models is safer than a few hundred researchers inside closed labs. He also says capability-based thresholds should govern the release of any model, open or closed. Stock argues open weights accelerate cyber defense and enable outside audits, and says Mistral's staged release buys extra testing time.

What this means

The pitch from both labs is control, not raw capability. Stock cites data and IP control, business continuity, customization and cost, and Laskin frames it as moving from renting intelligence to owning it. That targets governments and enterprises that avoid closed US labs for dependency reasons and avoid Chinese models for security reasons.

The claims are still unproven. Reflection's own benchmarks are mixed, and its efficiency figure is untested. Mistral has not published comparisons, and its "best open-weight model" claim cannot be checked until the weights ship. Independent evaluations after the October releases will show whether the Western gap has really narrowed.

The staged release is also a notable precedent. Mistral is withholding weights until after added reinforcement learning and safety testing, and giving partners a less restricted cyber-capable variant first. If that approach holds, open-weight releases at this scale may routinely trail API access by weeks.

Related Articles

model release

Thinking Machines Lab releases Inkling 2 Mini: 26B-parameter looped MoE with 3B active and 1M-token context

Thinking Machines Lab has released Inkling 2 Mini, a multimodal Mixture-of-Experts model with about 26B physical total parameters and 3B active per token. It repeats its middle layers three times, supports a 1M-token context, and ships with a DSpark drafter. Hugging Face reports it leads 32 of 48 benchmarks against four comparable models.

model release

Tencent releases Hy4 flagship weights: 770B-parameter MoE, 1,048,576-token context, conditional license

Tencent released downloadable weights for its flagship Hy4 on November 20, with a System Card. The multimodal MoE has a 770-billion-parameter language backbone, 49 billion active parameters per text token and a 1,048,576-token context window, under a license with commercial conditions for model-as-a-service operators.

model release

KAIST MLP Lab releases KORMo-Web-11B: open LLM, VLM and web agent scoring 30.1 on Korean Online-Mind2Web

KAIST MLP Lab has released KORMo-Web-11B, an 11B-parameter bilingual model that combines text, vision and browser-agent capabilities in one set of weights under Apache-2.0. On 612 Online-Mind2Web tasks it scores 29.0 in English and 30.1 in Korean, according to the lab.

model release

Tencent releases Hy4 flash: 220B MoE, 12B active parameters, 1M-token context

Tencent has released Hy4 flash, a multimodal Mixture-of-Experts model with 220B total parameters, 12B active per token, and a 1,048,576-token context window. Tencent calls it its most cost-effective Hy model to date and claims a per-task cost of 1/20 that of Hy4 on its internal WorkIndex benchmark.

Comments

Loading...