Standard Compute
← All models
Reflection · Balanced tier

Beam 1.0

Reflection's open-weight model for coding and agent work, released October 16, 2026: 501B parameters with 23B active, a 1M-token context and an Apache 2.0 license.

Beam 1.0 lists at $1 per million input tokens and $3 per million output tokens on OpenRouter as of October 17, 2026, with cache reads at $0.10. The four hosts there charge from $0.25 / $1 to $1.30 / $4.30 and serve it at different precisions. Reflection publishes no API price; its own platform is a free public beta. Source: OpenRouter: Beam 1.0

Reflection
Maker
64/100
Overall
57
Value score
$1.60
Blended /1M
1.0M
Context
Open
Weights
Pricing: $1.00 in · $3.00 out / 1M tokens · synced from OpenRouter · updated 2026-10-17

Scorecard

Output Quality7.0/10
Agentic Ability5.0/10
Speed7.0/10
Value for Money5.7/10
Reliability8.0/10

Beam 1.0 is Reflection's open-weight model, released October 16, 2026 under the Apache 2.0 license and built, Reflection says, for coding, reasoning and long-horizon agent work. It is a sparse mixture-of-experts model with 501B total and 23B active parameters (256 experts) and a 1M-token context. It reads and writes text only, was developed predominantly for English, and has a June 30, 2026 knowledge cutoff. The weights are on Hugging Face in BF16, FP8 and NVFP4 with no sign-up. Reflection reports 80.1% on Terminal-Bench 2.1 for the BF16 weights; that is Reflection's own run, and every benchmark figure published so far comes from Reflection. On OpenRouter it lists at $1 per million input and $3 per million output tokens, and Reflection's own platform is a public beta that is free to start. Standard Compute does not serve it yet.

Sources: Reflection's announcement · Hugging Face model card · Apache 2.0 license · OpenRouter listing

Strengths

  • Apache 2.0: commercial use, fine-tuning and redistribution with no revenue threshold or required product label
  • Open weights in BF16, FP8 and NVFP4, downloadable without a sign-up
  • $1/$3 per million on OpenRouter, with a 1M-token context

Trade-offs

  • Every published benchmark figure is Reflection's own so far
  • Text only: no image input
  • Hosts on OpenRouter differ in price ($0.25/$1 to $1.30/$4.30 per million) and precision (FP8 or FP4), so results can vary by host
  • Self-hosting needs a multi-GPU server: the weights are about 1 TB in BF16 and 294 GB in NVFP4
  • Not served on Standard Compute yet

Agents that support Beam 1.0

Trying a permissively licensed open-weight coding model through OpenRouter, or self-hosting it where code must stay on your own hardware.

Beam 1.0 FAQ

Is Beam 1.0 good for AI agents?
Trying a permissively licensed open-weight coding model through OpenRouter, or self-hosting it where code must stay on your own hardware. On our scoring, which is calibrated to public benchmarks, it rates 5.0/10 for agentic ability and 7.0/10 for output quality. Apache 2.0: commercial use, fine-tuning and redistribution with no revenue threshold or required product label.
How much does Beam 1.0 cost?
Beam 1.0 lists at $1 per million input tokens and $3 per million output tokens on OpenRouter as of October 17, 2026, with cache reads at $0.10. The four hosts there charge from $0.25 / $1 to $1.30 / $4.30 and serve it at different precisions. Reflection publishes no API price; its own platform is a free public beta. Via OpenRouter it is listed at $1.00 in · $3.00 out / 1M tokens.
What is Beam 1.0's context window?
Beam 1.0 has a 1.0M-token context window, large enough to hold sizeable codebases or document sets in a single run. Reflection is the maker.
Is Beam 1.0 open source?
Yes. Beam 1.0 ships with open weights, so you can self-host it or run it through a hosted provider such as OpenRouter. Check the license before commercial use.
What does Beam 1.0's Apache 2.0 license allow?
Apache 2.0 lets you use, modify, fine-tune, redistribute and sell the model and your derivatives, commercially or not, and it includes a patent grant. If you redistribute the weights or a derivative, include a copy of the license, keep the existing copyright and attribution notices, and mark files you changed. There is no revenue threshold, user cap or required "built with" label. Reflection's model card adds that it does not prohibit distillation, training other models on Beam's outputs, fine-tuning, quantization or merging, and points to Reflection's Usage Policy for areas to review; the card says it does not modify or restrict the license. This is a summary, not legal advice.
How do I use Beam 1.0 in a coding agent today?
The simplest way is OpenRouter, which lists it as reflection/beam-1.0-501b-a23b. In an agent that takes an OpenAI-compatible endpoint, such as OpenCode, Cline, Kilo Code, Hermes Agent or OpenClaw, set the base URL to https://openrouter.ai/api/v1, use your OpenRouter key and the model ID reflection/beam-1.0-501b-a23b. Reasoning is always on there and defaults to max effort; set reasoning_effort (low, medium, high, xhigh or max) to trade depth for speed and cost. The hosts there differ in price and precision, so check the provider list if that matters to you. Reflection also offers its own API at platform.reflection.ai, a public beta that is free to start, and Mirror, an open-source command-line agent harness for Beam.
Can I self-host Beam 1.0?
Yes, on a multi-GPU server. The weights download from Hugging Face without a sign-up in three precisions, about 1 TB in BF16, 503 GB in FP8 and 294 GB in NVFP4, plus a drafter model for speculative decoding. Reflection names Transformers, vLLM, SGLang, Unsloth, Axolotl, TokenSpeed and Dynamo as supported; vLLM and SGLang expose an OpenAI-compatible API your agent can point at. The model card gives no minimum hardware. It uses a modified Harmony chat template, and Reflection ran its evaluations at temperature 0.7 and top_p 0.99, the defaults in the repository's generation config.
Does Standard Compute serve Beam 1.0?
Not yet. Standard Compute's models are listed at standardcompute.com/models. On the models it serves, the router sends each step to the cheapest model that can do it, on providers vetted for price, uptime and speed with automatic fallback; opt-in Europe or US rules limit which provider endpoints may serve a key, and prompts are not used for training.
What are the downsides of Beam 1.0?
The main trade-offs: Every published benchmark figure is Reflection's own so far; Text only: no image input; Hosts on OpenRouter differ in price ($0.25/$1 to $1.30/$4.30 per million) and precision (FP8 or FP4), so results can vary by host; Self-hosting needs a multi-GPU server: the weights are about 1 TB in BF16 and 294 GB in NVFP4; Not served on Standard Compute yet. It's strongest for trying a permissively licensed open-weight coding model through OpenRouter, or self-hosting it where code must stay on your own hardware.

Beam 1.0 compared

All 26 Beam 1.0 comparisons

The agent matters as much as the model. Which agent gets the most out of Beam 1.0? →

Current Claude, GPT, Gemini and open models run on Standard Compute

Standard Compute

Current frontier models Claude Fable 5.1 · GPT-6 Astra

More intelligence
per dollar.

Each step goes to the cheapest model that can do it, on providers we vet for price, uptime and speed, with automatic fallback when one fails.

Compliance
Opt-in rules keep a company, team or key on European or US providers
Privacy
We don’t train on your prompts or store them

Fixed monthly plans from $19 · cancel anytime

7-day fair refund Free trial, no card Billing by Stripe3,500+ agent users