Reflection releases Beam's weights: a 501B open model at $0.25 per million input tokens
Reflection releases Beam, a 501B-parameter open-weight coding model under Apache 2.0, at $0.25 in and $1 out per million tokens on Nebius and free for now on Vercel. Elorian launches Glimpse, a visual reasoning model at $0.60/$3, Vercel and Shopify ship an agent-ready storefront template, and GitHub's license compliance checks go GA.
- Reflection released the weights of Beam, a 501B-parameter mixture-of-experts model (23B active) with a 1M-token context, under Apache 2.0; Nebius serves it at $0.25 in and $1 out per million tokens.
- On Reflection’s own comparison table, Beam trails DeepSeek V4.1 Flash, GLM 5.3 and Kimi K3 on Terminal-Bench 2.1, so its pitch is a cheap, Apache-licensed US model rather than the best open one.
- Elorian launched Glimpse, a model that zooms, crops and rewatches video while it reasons, at $0.60 in and $3 out per million tokens, 20% below Gemini 3.8 Flash’s introductory price.
- Vercel Shop, an open-source Next.js storefront for Shopify that agents can search and fill carts in over MCP, is generally available, along with Shopify’s rebuilt Hydrogen SDK.
- GitHub’s open source license compliance is GA for Enterprise Cloud, with an API that lets coding agents check a package’s license before adding it.
Beam is a cheap US open model, not the best open one
Reflection released Beam’s weights on Hugging Face on 16 October under Apache 2.0, eleven days after previewing it. It’s a text-only mixture-of-experts model, 501B parameters with 23B active, a 1M-token context and five reasoning levels, in BF16, FP8 and NVFP4 checkpoints. Hugging Face’s write-up puts NVFP4 at about 280 GB and recommends four B200s to serve it.
You don’t need the GPUs. Nebius lists it at $0.25 in and $1 out per million tokens, and Vercel’s AI Gateway serves it free for a limited launch period. AWS, Google Cloud, Azure, Cloudflare, Together and about a dozen other hosts are listed as launch partners. Reflection also open-sourced Mirror, a coding CLI it trained Beam alongside.
| Model | Price in / out ($ per M) | Terminal-Bench 2.1 | HLE, no tools | GPQA Diamond |
|---|---|---|---|---|
| Beam | $0.25 / $1.00 | 80.1 | 36.2 | 90.5 |
| DeepSeek V4.1 Flash | $0.30 / $1.20 | 90.6 | 39.1 | 90.9 |
| GLM 5.3 | $1.40 / $4.40 | 88.2 | 42.3 | 91.7 |
| Kimi K3 | $3.00 / $15.00 | 88.3 | 46.9 | 93.5 |
Scores from Reflection’s comparison table, which takes rivals’ numbers from Artificial Analysis and DataCurve. Prices are Nebius’s list prices for all four; DeepSeek’s own API charges $0.30/$1.20 at peak and half that off-peak. Highlighted values are the best in each column.
Cheap and Apache-licensed, but out-scored at its price. DeepSeek V4.1 Flash costs about the same (less off-peak on DeepSeek’s own API) and beats Beam on most rows of Reflection’s own table. Beam’s case rests on Reflection’s estimate that it needs 3 to 4 times less inference compute than GLM 5.2 for similar reasoning scores, based on active parameters and output tokens and not yet checked independently, and on being American. Expect it to win US buyers who won’t run Chinese weights; everyone else should benchmark it against DeepSeek first. Price comparisons are on our models table.
Elorian’s Glimpse looks twice before it answers
Elorian, a Palo Alto startup led by former Google DeepMind researcher Andrew Dai, launched Glimpse, its first model, priced at $0.60 per million input tokens and $3 per million output. Mid-reasoning it can crop an image, rewatch part of a video at a higher frame rate or run code to measure something. It’s available by sign-up through an API and a chat app; TechCrunch calls it a research preview.
Why it matters. Robotics teams, labs and media companies with large video archives pay frontier prices today for counting, captioning and event detection. Elorian says a few hundred GPUs and an open-weight base model were enough to build it.
Strong on charts and captions, behind on harder video. In Elorian’s own tests, Glimpse beats GPT-6 Astra ($10/$50) on chart reasoning (92.9 vs 92.1 on CharXiv-R), object counting on FSC-147 and short-video captioning, at 6% of the price. It trails Astra on dense captioning and temporal reasoning, and trails Astra and Gemini 3.8 Flash on Video-MME v2 (56.7 vs 64.1 and 59.6). Against Gemini 3.8 Flash’s introductory $0.75/$3.75 it’s 20% cheaper, so the fight is with Google’s Flash tier, not with the frontier.
Vercel and Shopify build the store for agents
Vercel made Vercel Shop generally available: an open-source Next.js storefront template for Shopify, built on Shopify’s Hydrogen SDK, which was rebuilt to work across JavaScript frameworks and is also GA. Each product URL serves a page to people and machine-readable price, stock and options to agents. The Universal Commerce Protocol, which Shopify co-developed, is exposed at the merchant’s domain, so agents can search and build carts over MCP.
Why it matters. Agent shopping today means scraping pages built for humans. This gives a Shopify merchant one codebase for the storefront, a built-in shopping assistant and the agent interface, with skills that guide coding agents through setup.
Checkout stays with Shopify. Vercel gets more Shopify storefronts on its hosting, and Shopify keeps checkout and payments while agents become the front end. That’s the distribution play: the protocol is open, but the template, the assistant and the deploy button all point at Vercel.
Quick hits
- GitHub’s open source license compliance is GA for Enterprise Cloud repos with Code Security: an enterprise-wide license policy blocks pull requests that add disallowed dependencies, and a new API lets coding agents check a package before suggesting it.
- Cloudflare will score Verified bots on behaviour: a “trust band” based on whether a bot follows robots.txt and slows down after HTTP 429 responses, used internally now, with public scores on Radar and customer-set thresholds planned.
- AWS Transform converts Azure SDK code to AWS, now GA for .NET apps: Azure Functions to Lambda, Blob Storage to S3 and Queue Storage to SQS, with Cosmos DB to DynamoDB in preview.
- Wiz’s runtime sensor becomes a managed GKE add-on in preview, with GKE handling deployment, credentials and updates for Standard and Autopilot clusters.