Reflection opens Beam, a 501B Apache-licensed coding model.
Reflection’s Beam opens a 501B coding model under Apache 2.0, with 2-bit builds small enough to try on a 256 GB Mac Studio
The license is the headline. Apache 2.0 is about as permissive as model licenses get, and Beam’s model card says Reflection does not prohibit distillation, training other models on its outputs, fine-tuning, quantizing, or merging. Compare that with last week’s Muse Spark Open, which requires a “Built with” credit and a separate deal for large model-service businesses. Reflection ships BF16, FP8, and NVFP4 checkpoints, a speculative-decoding drafter, its own open-source Mirror CLI coding harness, and a safety technical report.
The Mac angle is new too. Reflection’s NVFP4 build (about 280 GB) targets NVIDIA Blackwell GPUs, and Hugging Face suggests four B200s for real serving. But community GGUF files from Unsloth, one of Reflection’s named launch partners, list a 2-bit build at about 187 GB and a 3-bit build at about 242 GB. Unlike Muse, that puts a frontier-size open model under a 256 GB Mac Studio’s memory, at least on paper. Nobody has published Mac numbers, and macOS lets the GPU use less than all of physical memory by default, so 2-bit is the realistic try. Ollama lists Beam as cloud-only, and there is no MLX build yet.
Cautions. Every benchmark is Reflection-reported, and Reflection’s own announcement post concedes that Kimi K3 is ahead on raw capability; it sells efficiency instead, claiming 3–4× less inference compute than GLM-5.2 by its own estimate. Google Cloud’s “up to 5x higher throughput and 8x lower latency” is a vendor serving claim on GB300 racks. And nobody has measured what 2-bit does to Beam’s coding quality; Unsloth publishes divergence charts, not task scores.
What to do: run your own agent tasks through beam:cloud on Ollama or a hosted provider first. If it beats your current model, try the 2-bit GGUF locally on llama.cpp. If you are building a product or a fine-tune on an open model, the license alone earns Beam a slot in your eval set.
Sources: Reflection · Hugging Face blog · Model card · Unsloth GGUF files · Google Cloud
Anthropic says it will commit 5% of its shares, subject to stockholder approval, to a new Anthropic Public Benefit Fund, with up to 2% more at the board’s discretion for each 50% rise in the company’s value. Anthropic
NVIDIA posted a weekend roundup of RTX Spark launch events as the new Windows PCs reached buyers. NVIDIA blog
Model & runtime releases
| Item | Date | Badge |
|---|---|---|
| Reflection Beam 1.0 501B-A23B (BF16 / FP8 / NVFP4, Apache 2.0) | Oct 16 | Open |
| llama.cpp b11701 Beam support · b11714 Discover Models UI | Oct 16–17 | Open |
| Transformers v5.20.1 (Beam) | Oct 16 | Open |
| Nemotron-Labs Dory 1.0 4B looped model (OpenMDW) | Oct 16 | Open |
| Ollama v0.40.6 (Q2_0 GGUF) · v0.40.7-rc0 | Oct 16–17 | Open |
| GPT-Live prompt-to-voice + nine new voices | Oct 16 | Voice |
| DeepSeek V4.1-Flash-1017 (MIT weights) + Search & Fetch API | Oct 17 | Open |
| mlx-community: DeepSeek 1017 MLX quants · FLUX.3 Image Klein MLX | Oct 17–18 | Open |
| Tencent Hy-Memory-30B-A3B memory-writer model (Apache 2.0) | Oct 15–16 | Open |
| vLLM v0.32.0 | Oct 18 | Open |
| SGLang v0.5.22 (Apple Silicon PyTorch path) | Oct 18 | Open |
| JAX v0.12.0 | Oct 19 | Open |
Security / policy / access
| Item | Date | Badge |
|---|---|---|
| Beam license: Apache 2.0; distillation and output training allowed | Oct 16 | Active |
| Beam Safety Technical Report published | Oct 16 | Active |
| Cloudflare monitors Verified bots’ ongoing behavior | Oct 16 | Active |
| vLLM 0.32.0: default CORS closed + three advisory fixes | Oct 18 | Active |
| vLLM advisory ID mismatch (GHSA-4hhp) — track by PR | Oct 18 | Watch |
| SGLang: diffusion scheduler socket pinned to loopback | Oct 18 | Active |
| DeepSeek Search API: another place agent queries go | Oct 17 | Watch |
| llama.cpp gguf-split rejects truncated files | Oct 19 | Active |
| No newly published llama.cpp / Ollama / MLX / vLLM advisories | Oct 16–19 | Quiet |
Critical
Upgrading to vLLM 0.32.0? Browser clients break until you set allowed origins, and Python 3.10 is gone. Moving to JAX 0.12? Pre-Haswell x86 CPUs are no longer supported.
Watch
Reflection says Beam self-distillation recipes arrive “in the coming days.” No MLX build of Beam yet. Dory support is still in open pull requests for vLLM, llama.cpp, and mlx-lm. NVIDIA’s 4 GB Nemotron 3.5 Bolt is due Oct 23.
Opportunity
Nobody has published Beam numbers on a 256 GB Mac Studio yet. Reflection is giving $10,000 in Hugging Face credits to the best Beam projects shared on the Hub, and “quantize it for your favorite hardware” is on its list.
Run Beam at 2-bit on a 256 GB Studio
Update llama.cpp to b11701 or newer, then hf download unsloth/Beam-1.0-501B-A23B-GGUF --include "UD-Q2_K_XL/*" --local-dir beam (about 187 GB) and start llama-server -m beam/UD-Q2_K_XL/Beam-1.0-501B-A23B-UD-Q2_K_XL-00001-of-00005.gguf -c 32768. Watch memory pressure; if Metal refuses the allocation, cut the context first.
Same task, cloud versus local
ollama run beam:cloud (needs an Ollama account) with the exact prompt you gave the local build. Compare the answers and the wall-clock time. That tells you what 2-bit costs you.
Try SGLang’s new Mac path
In a Python 3.11+ environment, uv pip install --prerelease=allow sglang==0.5.22, then python -m sglang.launch_server --model-path Qwen/Qwen3-0.6B. Keep it to a small, unquantized text model; the Mac path rejects anything else at startup.
Friday through Monday morning (Oct 16–19 ET). ASTGL / Fraser takes, sourced.
Apache is the feature
The interesting thing about Beam isn’t the parameter count. It’s the license. A week ago Meta’s open model came with an attribution badge and a revenue clause. Beam comes with Apache 2.0 and a model card that says distillation and training on outputs are fine. If you are building a product or a fine-tune on an open model, that is a legal review you get to skip. My take: Beam earns a slot in your eval set on the license alone. Whether it keeps that slot depends on your tasks, because every score so far is Reflection-reported, and Reflection itself says Kimi K3 is ahead on raw capability.
Sources: Reflection · Model card
The 256 GB Studio gets a frontier-size candidate, at 2 bits
Muse needed more than 256 GB even at 3-bit. Beam’s 2-bit files come to about 187 GB. That leaves real room for context on a 256 GB Mac Studio, which is new for a model this size. Two catches: macOS gives the GPU a working budget smaller than physical memory, and nobody has published what 2-bit does to Beam’s coding. My take: this is the experiment of the week for Studio owners. Run the same three agent tasks through the cloud build and the local 2-bit build and write the numbers down. If 2-bit holds up, you have a frontier-size coder with no API bill.
Sources: Unsloth GGUF files · llama.cpp #30670 · SGLang #36780 (Metal memory budget)
DeepSeek’s quiet upgrade is really a search API
Over the weekend DeepSeek pointed
deepseek-flashat a new V4.1 Flash build at the same price, with benchmark gains that are DeepSeek-reported. The weights are on Hugging Face under MIT, but at about 510 GB they are a datacenter download; even the community 3-bit MLX build is about 380 GB. The bigger change for operators is a first-party Search and Fetch API: web results, or any page as Markdown, on the same API key. My take: a model vendor selling grounding as a plain API is a cheap way to give agents fresh web context. It is also one more place your agents’ queries go. Read the data terms before you wire it into anything that touches customer data.Sources: DeepSeek change log (Oct 17) · Search & Fetch guide · Hugging Face
vLLM finally closes the door on CORS
vLLM 0.32.0 landed early Sunday, and the line that will bite people is under breaking changes: the default CORS setting no longer allows every origin. If a browser app talks straight to your vLLM server, it breaks until you set allowed origins. The release also lists three security fixes, signs shared-memory handles, rejects absurdly nested JSON schemas, and drops Python 3.10. My take: this is the right default and long overdue. Upgrade, then list your origins on purpose instead of putting the wildcard back. One wrinkle: one advisory ID in the notes resolves to a different, older bug on GitHub, so track these fixes by pull request until the advisories publish.
Sources: vLLM v0.32.0 · GHSA-4hhp-h66f-j5j7
SGLang shows up on the Mac, with training wheels
SGLang 0.5.22, out Sunday night, runs on Apple Silicon through its standard PyTorch path, beside the MLX backend it already had. It is deliberately narrow: no quantized weights, no images, no speculative decoding, one device. My take: this is a dev-parity win, not a Mac inference server. If your production stack is SGLang on NVIDIA, you can now step through the same scheduler code on your laptop. For actually serving models on a Studio, MLX and llama.cpp are still the tools.
Sources: SGLang v0.5.22 · PR #36780