Reflection Beam and Mistral Large 4: Two Open-Weight Heavyweights, No Weights Yet

Two large open-weight model launches landed in roughly 48 hours. Reflection AI introduced Beam on October 5, 2026. Mistral launched a public preview of Mistral Large 4 (ML4, "Le Chonk") on October 6. Neither set of weights is downloadable today.

That is the story for builders. You can join a waitlist or hit a guarded API. You cannot drop either checkpoint into your own stack yet. Vendor numbers are either self-reported (Reflection) or still being filled in (Mistral). Treat this week as announcement week, not migration week.

Reflection Beam: sparse MoE, waitlist now, weights later this month

Reflection's introducing Beam post describes a sparse Mixture-of-Experts model with 501 billion total parameters and 23 billion active, aimed at coding, reasoning, and agent workloads. It is text-only. Midtraining extends effective context to 1M tokens. Users get a reasoning-effort control that trades shorter answers for lower compute or longer reasoning for harder tasks.

Access today is early-access / waitlist while red-teaming finishes. Reflection says it will release the weights, technical report, model card, and developer artifacts later this month under Apache 2.0.

Reflection is blunt about the competitive frame. It says frontier open models like Kimi K3 remain ahead on raw capability, and pitches Beam on inference efficiency: competitive coding and agent scores with less compute per token than some larger open models. The benchmark tables on that page are Reflection's own results. One example they report: SWE-Bench Verified at 80.9. Do not treat that as an independent leaderboard until third parties run the weights.

Training detail that matters for credibility, not for your VRAM bill yet: Reflection says Beam was pretrained on 23.8 trillion tokens, then pushed through a high-compute RL run on 10.5K NVIDIA GB300 GPUs generating more than 100 million rollouts over about four weeks. Useful context for how they built it. Irrelevant until you can actually download the checkpoint.

Mistral Large 4: ~1T multimodal, API preview, weights end of month

Mistral's own Large 4 announcement calls ML4 a 1 trillion-parameter natively multimodal model with 49 billion active parameters. Nickname on the page: Le Chonk. You can try a public preview API on Mistral Studio now. Weights are not out.

On timing, use Mistral's words: "Weights drop end of this month" and "We will release the weights by the end of the month," after red-teaming with cybersecurity partners, vetted partners, and state authorities. TechCrunch's coverage paraphrases that as about three weeks after safety testing. We are not stating a calendar day that Mistral did not publish on its own page.

Mistral says ML4 was trained from scratch on 3,800 NVIDIA Grace Blackwell GPUs in its own European datacenters, and that the preview is served on that same stack. The company highlights cybersecurity, finance, and related enterprise workloads. It also says more architecture detail, additional benchmarks, and post-training notes will come as weights approach. So if a secondary write-up says "benchmarks pending," that matches Mistral's own "more to come" language even though the announcement already shows some early vendor charts.

TechCrunch's Oct 6 piece adds color from Mistral VP Science Pierre Stock on open weights as easier to audit, and on keeping the model available to defenders without handing it to attackers early. That is positioning. It is not a substitute for your own evals.

Same week, different shapes

Beam (Reflection): announced Oct 5, 2026. 501B total / 23B active MoE. Text-only. Waitlist now. Apache 2.0 weights later this month. Self-reported benches on Reflection's post.

Mistral Large 4: announced Oct 6, 2026. About 1T / 49B active MoE. Multimodal. Guarded preview API on Mistral Studio. Weights end of this month per Mistral. Early vendor charts, with more detail promised alongside weights.

Both are MoE. Both sell "open weight soon." Neither is a drop-in base model for your cluster tonight.

What to check before you swap an open base model

Once weights actually ship, the questions that matter for a coding or agent stack are boring and specific:

1. License and redistribution. Beam is promised under Apache 2.0. Confirm the Mistral weight license on the release day, not from preview marketing.
2. Active vs total parameters. Hosting cost tracks active params and serving shape more than the headline total. Beam's pitch is 23B active. ML4's announcement puts active at 49B. Measure tokens/sec and VRAM on your hardware, not on a press table.
3. Context and tool use. Beam advertises 1M effective context and a reasoning-effort knob. ML4 is multimodal and agent-oriented in Mistral's demos. Match the model to whether your workload is long-repo text, tool calling, or vision-heavy docs.
4. Independent evals. Re-run SWE-bench / Terminal-Bench / your private golden set after download. Vendor DeepSWE or cyber index numbers are starting points, not acceptance tests.
5. Safety and refusal profile. Mistral is spending weeks on cyber red-teaming before weights. Reflection is still red-teaming Beam. If you run security or customer-data agents, test jailbreaks and over-refusal on your own prompts before you cut over.
6. Swap cost. Changing a base model means redoing adapters, prompts, evaluators, and sometimes the tool schema. Do not schedule that on announcement day.

We are watching both releases the same way we watch any open base cutover: wait for the files, read the license, run our own benches, then decide.

What we are not claiming

No hosting dollar figures. No invented ship dates. No claim that either model beats a closed frontier system on an independent suite. Reflection's efficiency story and Mistral's vertical charts are theirs until weights and third-party runs exist.

Sources

• Introducing Beam: Reflection's 501B open-weight model (Reflection AI, Oct 5, 2026)
• Reflection AI debuts open-source Beam model with 501B parameters (SiliconANGLE, Oct 5, 2026)
• Introducing Mistral Large 4 (Mistral, Oct 6, 2026)
• Mistral Large 4 model docs (Mistral Docs, public preview)
• Mistral's new 1T model aims to leapfrog closed and open rivals (TechCrunch, Oct 6, 2026)