Reflection Ships Beam’s 501B Open Weights Under Apache 2.0 With API Prices From $0.25 per Million Tokens, but the Smallest Official Build Is Still 294 GB

Reflection AI has put the weights for Beam, its 501 billion parameter coding and agent model, on Hugging Face under an Apache 2.0 license. The company announced the release on 16 October 2026 in a post titled “Beam is now available”, alongside a final model card and a full safety report.

That closes the gap we flagged when Reflection opened the Beam waitlist on 5 October with no weights and no price. Now there are both. Hosted prices on OpenRouter start at $0.25 per million input tokens, and the release became a trending story on X on Friday.

The catch is size. Even the most compressed official build is roughly 294 GB of weights, so “open” here means open to anyone with a serious GPU cluster or a cloud account, not a laptop.

What Reflection Actually Released

According to the Beam model card, Beam is a sparse Mixture of Experts model with 501.3B total parameters and 23.4B active per token. It has 52 layers (3 dense and 49 MoE), 256 experts, a 152k token vocabulary and a 1M token context window. It is text only, was developed predominantly for English, and has a knowledge cutoff of 30 June 2026.

Reflection shipped four repositories. We added up the safetensors file sizes listed on each Hugging Face repo to see what you would actually download.

Hugging Face repo	Format	Weights on disk (our sum)
Beam-1.0-501B-A23B	BF16	about 1,003 GB
Beam-1.0-501B-A23B-FP8	FP8	about 503 GB
Beam-1.0-501B-A23B-NVFP4	NVFP4 (4-bit)	about 294 GB
Beam-1.0-501B-A23B-DFlash2	Speculative decoding drafter	about 3.9 GB

Those figures cover weights only. Long context needs extra memory for the KV cache on top, and a 1M token window is not free. NVFP4 is NVIDIA’s 4-bit format built for its Blackwell GPUs, which is the hardware Reflection says it serves Beam on in production.

The license matters as much as the files. The model card says Reflection does not prohibit distillation, using outputs to train other models, fine-tuning, quantization or merging. For startups that want to build on a model rather than just call it, that matters as much as any benchmark.

Final Benchmarks Moved, a Little, Since the Preview

Every benchmark below is still Reflection’s own. What is new is that we can compare the final model card against the numbers Reflection posted in its launch post, which it last updated on 8 October.

Benchmark (Reflection-reported)	Preview (8 Oct)	Final BF16 (16 Oct)	Final NVFP4
SWE-bench Verified	80.9%	81.6%	80.7%
Terminal Bench v2.1	80.1%	80.1%	78.4%
DeepSWE v1.1	44.4%	43.3%	42.7%
SWE Bench Pro v1	65.5%	66.1%	65.6%
HLE (no tools)	36.2%	36.0%	34.8%
GPQA Diamond	90.5%	90.6%	89.3%
MCP Atlas	78.7%	80.8%	80.5%

Two things stand out. First, the 4-bit build gives up less than two points on every test in this table, which makes the 294 GB version the realistic choice for most self-hosters. Second, Beam is still well behind some Chinese open models on the hardest agent coding test. In Reflection’s own preview table, Kimi K3 scored 68.0% on DeepSWE and DeepSeek V4.1 Flash scored 74.2%, against Beam’s 43 to 44%.

Reflection’s pitch is efficiency rather than raw score. The model card claims Beam reaches comparable quality to leading open models with roughly 3x fewer reasoning tokens. That is a vendor claim, but it is the one claim an outside lab has now partly tested.

What Artificial Analysis Measured

Independent benchmarker Artificial Analysis scores Beam 1.0 at 30 on its Intelligence Index, 19th of the 120 models it ranks, against a median of 19 for comparable models. It measured output at about 228 tokens per second, second fastest of 77 models tracked.

The token efficiency claim holds up in its data. Beam used 79M output tokens to run the full index, compared with a median of 130M. Artificial Analysis still calls the model “somewhat expensive” next to open models of similar size, at $0.70 input and $2.10 output per million tokens versus medians of $0.43 and $1.68. Its cost to run one index task works out to $0.57.

Worth noting: Artificial Analysis is listed among Reflection’s pre-release testing partners in the launch post. That does not make its numbers wrong, but it is the kind of relationship readers should know about.

What It Costs to Rent Beam Today

If you do not want to touch 294 GB of weights, OpenRouter already lists four providers. Prices vary by about 5x, mostly because providers run different precisions.

Provider (via OpenRouter)	Precision	Input per 1M tokens	Output per 1M tokens
Nebius	FP8	$0.25 (about ₹24)	$1.00 (about ₹95)
Together	Not stated	$0.40 (about ₹38)	$1.20 (about ₹114)
CoreWeave	FP4	$1.00 (about ₹95)	$3.00 (about ₹285)
Baseten	FP4	$1.30 (about ₹124)	$4.30 (about ₹409)

Rupee figures use about ₹95.1 to the US dollar on 17 October 2026. Prices were read from OpenRouter’s public endpoint data the same day and can change without notice. Reflection’s own API is free to start during a public beta, and we could not find a published per-token rate for it.

For context, Google’s Gemini 4 Argon launched at $2 per million input tokens. Beam is not trying to beat it on capability. It is pitching a cheaper model you can also download and keep.

The Safety Report Says “Limited” Risk, With Caveats

The Beam Safety Technical Report rates the model as Limited risk across biological weapons, chemical weapons, loss of control and cyber offense under the Open Safety Framework. Reflection defines Limited as no material shift in the offense and defense balance compared with other open models.

Some of the detail is less comfortable. On SecureBio’s Virology Capabilities Test, Beam scored 52% against a human expert reference of 21.1%, tied for third of 13 models tested. Reflection’s argument is that rival open models already score in the same range, so releasing Beam adds little new risk.

The prompt injection numbers are more useful for builders. With no defense prompt, Beam carried out a planted malicious instruction in 3.10% of attacked episodes on Reflection’s MirrorDojo test and 3.60% on MirrorDyn. With the short anti-injection policy that ships in Reflection’s Mirror harness, that fell to 0.3% and 0%. Reflection itself warns these results are not directly comparable to public AgentDojo scores, because it ran every model inside its own harness.

On Reflection’s in-house safety suite, Beam refused or redirected 88.9% of harmful requests and complied with 81.5% of harmless ones. Put another way, nearly one in five harmless requests in that test did not get a straight answer.

Confirmed vs Not Yet Proven

Claim	Status
Weights on Hugging Face under Apache 2.0, in BF16, FP8 and NVFP4	Confirmed on the Hugging Face repos
501B total and 23B active parameters, 1M context	Confirmed in the model card
Hosted access from $0.25 input per million tokens	Confirmed on OpenRouter (Nebius, FP8)
Intelligence Index 30 and about 228 tokens per second	Measured by Artificial Analysis, a launch testing partner
SWE-bench Verified 81.6% and other coding scores	Reflection-reported, not independently reproduced
Roughly 3x fewer reasoning tokens than leading open models	Partly supported by Artificial Analysis token counts
Self-distillation recipes	Promised “in the coming days”, not released yet

What It Means for Indian Developers

For most Indian teams, Beam is an API decision first. At $0.25 input and $1.00 output per million tokens through Nebius, it is cheap enough to test as a coding or agent backend, and the 1M context helps with large repositories. Run your own evals before switching, especially on Indian English, Hinglish and local codebases, since Reflection says the model was built mostly for English.

Self-hosting is a different story. The 503 GB FP8 build fits on a single node of eight 80 GB GPUs with limited room left for long context, and NVFP4 needs Blackwell-class hardware to make sense. Like Meta’s Muse Spark Open earlier this week, this is open weights for companies with data center budgets or strict data residency needs.

The permissive license is the real opening. Startups can fine-tune Beam, distill it into something smaller, and ship it inside products under a standard permissive license with no extra field-of-use clauses. That is worth more to an Indian SaaS team than any single benchmark row.

FAQ

Is Reflection’s Beam really open source?

The weights are released under Apache 2.0 on Hugging Face, and Reflection says it does not prohibit distillation, fine-tuning or training on outputs. Reflection has not released its training data, so it is open weights rather than fully open source.

How much memory does Beam need to run locally?

The weights alone are about 1,003 GB in BF16, 503 GB in FP8 and 294 GB in NVFP4, based on the file sizes on Hugging Face. You need more than that for the KV cache, especially at long context.

How much does the Beam API cost?

On OpenRouter, providers charge from $0.25 to $1.30 per million input tokens and $1.00 to $4.30 per million output tokens as of 17 October 2026. Reflection’s own API is free to start during a public beta.

How good is Beam at coding?

Reflection reports 81.6% on SWE-bench Verified and 80.1% on Terminal Bench v2.1. It trails Kimi K3 and DeepSeek V4.1 Flash on DeepSWE in Reflection’s own comparison, and none of the coding scores has been independently reproduced yet.

Is Beam safe to use in agents?

Reflection rates it Limited risk and reports a 3.1% to 3.6% prompt injection success rate without defenses, falling near zero with its Mirror harness policy. Treat those as vendor figures and add your own tool permission checks.