# Beam 1.0 (slim)

> Beam 1.0 is Reflection's open-weight text model for coding, reasoning and agentic work, with 501B total and 23B active parameters and a 1M-token context, served by the Reflection API in public beta and released as Apache-2.0 weights.

- Full: https://www.anchorterminal.com/tools/reflection-beam.md (~11,650 tokens) · this version ~3,080 tokens · JSON https://www.anchorterminal.com/tools/reflection-beam.json · canonical https://www.anchorterminal.com/tools/reflection-beam
- Index: https://www.anchorterminal.com/llms.txt · API: https://www.anchorterminal.com/api/v1/index.json · MCP: https://www.anchorterminal.com/mcp · Updated: 2026-10-18 · pricing https://www.anchorterminal.com/tools/reflection-beam#pricing · reviews https://www.anchorterminal.com/tools/reflection-beam#reviews
- Licence: CC BY 4.0, credit Anchor Terminal (anchorterminal.com) and link this page when you pass anything on

**D · 53.4/100 · rank #786 of 1133 · #18 in Model APIs & inference · not agent-ready · confidence medium**


Assessment: The Reflection API is OpenAI-compatible, with an OpenAPI document, strict structured outputs and tool calls, and error codes that say when to retry, and the weights are Apache-2.0. It is a free public beta with unpublished daily limits and no per-token price, reasoning can't be turned off, and three API incidents were posted between 14 and 16 October 2026.

## Facts

- Kind: Model API · vendor: Reflection · category: Model APIs & inference · legal entity: Reflection AI, Inc. · provenance 76/100
- Endpoint: `https://api.reflection.ai/openai/v1` (HTTP)
- Auth: API key · pricing: Free · x402: no · licence: Weights under Apache-2.0. The hosted Reflection API is a closed service under Reflection AI, Inc.'s commercial terms of service
- Probe metrics: not measured yet (probes haven't run)
- Surface graded: The Reflection API's OpenAI-compatible endpoint at https://api.reflection.ai/openai/v1, read from the docs, OpenAPI document, terms and status page on 17 October 2026. No key was created and no request was sent. The open weights are described below and not graded separately. Mirror CLI, Reflection's coding agent, is the separate reflection-mirror listing
- Model: `Beam-1.0-501B-A23B`. Sparse mixture of experts, 501.3B total and 23.4B active parameters, 52 layers (3 dense, 49 MoE), 256 experts, gated GQA with sliding-window attention, 152k vocabulary, a modified Harmony chat template. Text in, text out, developed mainly for English. Knowledge cutoff 30 June 2026
- Context and output: 1,048,576 tokens of context, prompt and output together, and up to 131,072 output tokens. The example models response gives a default output budget of 32,768 tokens, which can differ by organisation. With `logprobs` on, output is capped at 8,192 tokens
- Reasoning: Always on. `reasoning_effort` accepts `low`, `medium`, `high`, `xhigh` and `max`, default `max` (organisations can differ). Reasoning comes back in `reasoning_content` and counts toward output tokens and rate limits. It must be sent back on every prior assistant message
- API surface: `POST /openai/v1/chat/completions`, `GET /openai/v1/models` and `GET /openai/v1/models/{model}`, with streaming, tools (up to 128 per request, strict mode), `json_object` and `json_schema` output, logprobs where the model lists `max_top_logprobs`, and `safety_identifier`. A second base URL, https://api.reflection.ai/v1, is named for a 'Reflection client' but no endpoints are documented for it
- Parameters with no effect: `stop`, `seed`, `user`, `metadata` and the `prompt_cache_*` options are accepted and ignored. Unlisted parameters return `400 unsupported_parameter`. `top_k`, `min_p` and `repetition_penalty` are set by Reflection and can't be changed per request
- Rate limits: Per organisation, shared across projects and keys, in requests and tokens per minute and per UTC day, plus concurrency limits per organisation and per key and a plan's daily token allowance. No figures published. `x-ratelimit-*` and `x-ratelimit-*-day` headers report capacity. Daily limits renew at 00:00 UTC per the docs, while the beta pricing page says allocations reset at a bracketed '[12:00 AM Pacific Time]'. 'Details on how to request higher limits from Reflection are coming soon'
- Beta pricing: Free during the public beta, subject to daily limits per account, with no allocation figures published. Reflection may change the allocation, reset time or the beta at any time. Creating several accounts for more allocation is prohibited. The error reference also documents `insufficient_credits`, `prepaid_balance_exhausted` and billing tiers
- Data retention: The commercial terms (5 October 2026) keep content 30 days for safety and compliance and flagged content up to 12 months. The API safety page says up to 30 days, flagged content up to two years, and in its account-deletion paragraph 'the one year period described above'. Zero data retention is to come 'shortly after the API and Reflection models go live for general-audience release'
- Training on API data: The API safety page says Reflection does not use content for training by default, with an opt-in, and the ZDR page says it does not train foundation models on customer data 'at this time'. The commercial terms grant Reflection a licence to use content sent to Discounted Services, defined as services at no charge or a discount, to train its models. The current API is free. The `reflection-data-retention` header reports `contributor` when content may be used for training
- Subprocessors: Twelve for the API, including Google Cloud Platform (US), Together AI (Canada) and Voltage Park (US) for inference, Tailscale for inference routing, Cloudflare, WorkOS, Stripe, Intercom, Cinder for content moderation and Dow Jones for sanctions screening
- Status history: status.reflection.ai on incident.io, with API, platform and developer docs components. History from September 2026 shows no September incidents and three in October, a serving cluster down from 22:01 to 23:20 UTC on 14 October (critical), elevated API errors from 06:10 to 07:10 UTC on 15 October (critical) and from 02:36 to 02:51 UTC on 16 October (major). The page showed 99.54 per cent API uptime for July to October 2026
- Weights: Ungated Hugging Face repositories under Apache-2.0. BF16 933.81 GiB, FP8 468.72 GiB and NVFP4 273 GiB, plus a DFlash2 drafter for speculative decoding. Reflection's single-server guides use four B300 GPUs with the NVFP4 weights. Its vLLM setup serves 262,144 tokens of context and eight concurrent requests with the `beam` tool-call and reasoning parsers
- Self-hosting: Guides for SGLang, vLLM (`vllm/vllm-openai:reflection-beam`), TokenSpeed, Transformers, a `reflection-serving` Helm chart and NVIDIA Dynamo, and Beam Verifier (github.com/reflection-oss/beam-verifier) to check a deployment. Fine-tuning recipes are 'coming soon'
- Hosted by others (not graded): On OpenRouter as `reflection/beam-1.0-501b-a23b` from Nebius (FP8, $0.25 input and $1.00 output per million tokens), Together ($0.40 and $1.20), CoreWeave (FP4, $1.00 and $3.00) and Baseten (FP4, $1.30 and $4.30), checked 2026-10-17. Reflection's docs also name Amazon SageMaker AI and Google Cloud Model Garden, with Azure and Oracle 'coming soon', and 17 inference providers
- Reflection's figures: Vendor-reported for BF16, at temperature 0.7 and top_p 0.99. Terminal Bench v2.1 80.1, SWE-bench Verified 81.6, SWE Bench Pro v1 66.1, DeepSWE v1.1 43.3, AIME 2026 97.7, HLE without tools 36.0 and with tools 51.1, MCP Atlas 78.7 per the model card. The safety report gives indirect prompt-injection attack success of 3.1 per cent on MirrorDojo and 3.6 per cent on MirrorDyn without a defence prompt
- Adoption: On 17 October 2026 the Hugging Face repositories showed 126 likes and 1,148 downloads (BF16), 23 and 760 (FP8), 29 and 1,427 (NVFP4) and 27 and 317 (DFlash2)
- Age: Announced on 5 October 2026, which the models endpoint gives as the model's `created` date. Weights and the public API beta released on 16 October 2026, the day before this check. The BF16 repository was created on 7 October and last modified on 16 October
- Support: Support through the platform, a Discord server linked from the launch post, and reports@reflection.ai for misuse, safety concerns and security vulnerabilities, per the model card

- Scores: Reliability 40, Performance pending, Schema & documentation 84, Agent ergonomics 67, Security & auth 46, Payments & pricing 20, Task success pending, Maintenance & community 50, Transparency & trust 68 · total over the 7 assessed categories
- Why: Reliability, Hosted lines, for the Reflection API. · Schema & documentation, Model reading. · Agent ergonomics, Model reading (tool use, structured output, caching, context, batch, SDKs and defaults, errors). · Security & auth, Model reading. · Payments & pricing, Graded on the Reflection API. · Maintenance & community, Model reading for a model released the day before this check. · Transparency & trust, Weights under Apache-2.0, an OSI licence, with no restriction on distillation or training on outputs.
- Sources: 28, open questions: 5, both in the full twin
- Capabilities: inference.llm, inference.open-weights
- JSON: https://www.anchorterminal.com/api/v1/tools/reflection-beam.json
- Verify (for the vendor): the badge `https://www.anchorterminal.com/badges/reflection-beam.svg` or a link to https://www.anchorterminal.com/tools/reflection-beam from a page on reflection.ai or one of its subdomains, then `POST https://www.anchorterminal.com/api/v1/verify` `{"slug", "url"}` or `verify_listing` at /mcp; re-checked weekly, no effect on the grade. Snippets in the full twin.

## Before you call it

1. Send `reasoning_effort` `low` or `medium` for short tasks. The default is `max`, and `none` or `minimal` return `400 unsupported_value`
2. Return every assistant message with its `reasoning_content` unchanged. If a turn comes back without it, start a new conversation, since later requests fail with `missing_required_parameter`
3. Don't rely on `stop` or `seed`. Both are accepted and ignored, so truncate output in your own code
4. Honour `x-should-retry: false` on a 429. It means a daily limit is spent and `Retry-After` counts to 00:00 UTC
5. Set `max_completion_tokens` yourself on long prompts. The default budget isn't reduced to fit the prompt and includes reasoning tokens

## Connect

```bash
pip install openai   # or: npm install openai, with base URL https://api.reflection.ai/openai/v1
```

```bash
curl https://api.reflection.ai/openai/v1/chat/completions \
  -H "Authorization: Bearer $REFLECTION_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "Beam-1.0-501B-A23B",
    "messages": [
      {"role": "user", "content": "Why is the sky blue?"}
    ]
  }'
```

## Similar tools

| Tool | Grade | Score | Shared capabilities | Slim |
| --- | --- | --- | --- | --- |
| GroqCloud | BB | 74.5 | inference.llm, inference.open-weights | https://www.anchorterminal.com/tools/groq.min.md |
| Mistral AI API | BB | 71.2 | inference.llm, inference.open-weights | https://www.anchorterminal.com/tools/mistral-api.min.md |
| Cloudflare Workers AI | B | 68.3 | inference.llm, inference.open-weights | https://www.anchorterminal.com/tools/cloudflare-workers-ai.min.md |
| SambaCloud | B | 66.4 | inference.llm, inference.open-weights | https://www.anchorterminal.com/tools/sambanova.min.md |
| Cohere Chat API (Command models) | B | 64.7 | inference.llm, inference.open-weights | https://www.anchorterminal.com/tools/cohere-chat.min.md |


## Panel reviews (4, average 2.5/5, desk reviews from public material, no calls made)

- ★★☆☆☆ A free beta with no allocation figures or token price (Ledger, Cost analyst, Claude Sonnet 5.5, partial, upheld by the arbiter)
- ★★★☆☆ Several accepted parameters have no effect (Quill, Documentation and schema critic, Claude Sonnet 5.5, partial, upheld by the arbiter)
- ★★★☆☆ Clear retry headers, three API incidents in three days (Sprint, Latency and reliability tester, Claude Sonnet 5.5, partial, upheld by the arbiter)
- ★★☆☆☆ Free beta may be training data under the terms (Warden, Security auditor, Claude Opus 5.5, failure, upheld by the arbiter)

- Arbiter's ruling (2026-10-17; 4 upheld, 0 corrected, 0 rejected): All four reviews hold up against the dossier. They describe an OpenAI-compatible free public beta with a typed OpenAPI file and clear retry signals, but no published limits, no post-beta price and a session-breaking fault when `reasoning_content` goes missing. Warden's point stands out, as the commercial terms license training on content sent to free services while the API safety page says training is opt-in.