# If the user wants more details, tell them they can access this page directly via the URL: https://hacksnap.live/story/beam-reflections-501b-open-weight-model-49969183

# Beam: Reflection's 501B open\-weight model

532 points · 166 comments

[Full discussion](<https://news.ycombinator.com/item?id=49969183>)

[Read original](<https://reflection.ai/blog/introducing-beam>)

Category: [Models & Products](<https://hacksnap.live/?category=models-products>)

## Skept-o-meter & Hotness

Skept\-o\-meter: High\. Estimated from 10 comments\.

2 comments for the summary\.

Peak rank: \#1

Time in Top 10: 24\.0 hours

Hacksnap ranks recent stories first, then orders each group by points\. Peak rank uses all retained observations\. Time in the Top 10 is estimated by holding each recorded rank until the next observation; gaps over 13 hours and time after the last observation are excluded\. Movement between observations is unknown\.

83 recorded rank observations from 2026\-10\-05T21:01:47\.650671\+00:00 to 2026\-10\-10T23:01:00\.95027\+00:00\.

Hotness — latest 83 recorded Hacksnap ranks:

2026\-10\-05T21:01:47\.650671\+00:00: rank \#6

2026\-10\-05T22:02:59\.93099\+00:00: rank \#3

2026\-10\-06T08:02:11\.210914\+00:00: rank \#2

2026\-10\-06T08:02:14\.354697\+00:00: rank \#2

2026\-10\-06T09:01:21\.688765\+00:00: rank \#2

2026\-10\-06T10:00:58\.047773\+00:00: rank \#2

2026\-10\-06T11:00:41\.689929\+00:00: rank \#2

2026\-10\-06T12:00:41\.7705\+00:00: rank \#1

2026\-10\-06T13:01:24\.509519\+00:00: rank \#1

2026\-10\-06T14:00:26\.950817\+00:00: rank \#1

2026\-10\-06T15:00:49\.066414\+00:00: rank \#1

2026\-10\-06T16:00:58\.166747\+00:00: rank \#2

2026\-10\-06T17:01:17\.009511\+00:00: rank \#2

2026\-10\-06T18:01:43\.791429\+00:00: rank \#2

2026\-10\-06T19:00:46\.048609\+00:00: rank \#2

2026\-10\-06T20:01:14\.687294\+00:00: rank \#2

2026\-10\-06T21:02:13\.414382\+00:00: rank \#52

2026\-10\-06T22:00:40\.926271\+00:00: rank \#50

2026\-10\-06T23:03:26\.867445\+00:00: rank \#51

2026\-10\-07T08:01:41\.339804\+00:00: rank \#52

2026\-10\-07T09:02:38\.026986\+00:00: rank \#53

2026\-10\-07T10:01:35\.069531\+00:00: rank \#53

2026\-10\-07T11:02:07\.770084\+00:00: rank \#53

2026\-10\-07T12:01:14\.996697\+00:00: rank \#52

2026\-10\-07T13:00:47\.45912\+00:00: rank \#51

2026\-10\-07T14:01:18\.307739\+00:00: rank \#51

2026\-10\-07T15:01:57\.541362\+00:00: rank \#53

2026\-10\-07T16:01:28\.474022\+00:00: rank \#52

2026\-10\-07T17:01:15\.036908\+00:00: rank \#53

2026\-10\-07T18:00:53\.300079\+00:00: rank \#52

2026\-10\-07T19:01:49\.456657\+00:00: rank \#53

2026\-10\-07T20:01:59\.032487\+00:00: rank \#53

2026\-10\-07T21:05:36\.655392\+00:00: rank \#55

2026\-10\-07T22:02:27\.105241\+00:00: rank \#56

2026\-10\-07T23:01:17\.452894\+00:00: rank \#55

2026\-10\-08T08:03:55\.083808\+00:00: rank \#55

2026\-10\-08T09:03:58\.68898\+00:00: rank \#53

2026\-10\-08T10:04:05\.364235\+00:00: rank \#54

2026\-10\-08T11:02:04\.410307\+00:00: rank \#53

2026\-10\-08T12:02:36\.39618\+00:00: rank \#53

2026\-10\-08T13:03:09\.206401\+00:00: rank \#53

2026\-10\-08T14:03:30\.113384\+00:00: rank \#54

2026\-10\-08T15:02:54\.444758\+00:00: rank \#52

2026\-10\-08T16:03:13\.067807\+00:00: rank \#51

2026\-10\-08T17:03:47\.92465\+00:00: rank \#50

2026\-10\-08T18:01:50\.360431\+00:00: rank \#50

2026\-10\-08T19:02:55\.719274\+00:00: rank \#51

2026\-10\-08T20:01:36\.359022\+00:00: rank \#51

2026\-10\-08T21:03:34\.000719\+00:00: rank \#51

2026\-10\-08T22:01:13\.459732\+00:00: rank \#50

2026\-10\-08T23:01:25\.318392\+00:00: rank \#49

2026\-10\-09T08:01:53\.314021\+00:00: rank \#50

2026\-10\-09T09:02:59\.93426\+00:00: rank \#51

2026\-10\-09T10:02:22\.652527\+00:00: rank \#50

2026\-10\-09T11:02:04\.807282\+00:00: rank \#50

2026\-10\-09T12:02:16\.769295\+00:00: rank \#50

2026\-10\-09T13:02:30\.35034\+00:00: rank \#52

2026\-10\-09T14:01:34\.215134\+00:00: rank \#51

2026\-10\-09T15:01:27\.564392\+00:00: rank \#50

2026\-10\-09T16:02:06\.561056\+00:00: rank \#51

2026\-10\-09T17:03:06\.673648\+00:00: rank \#53

2026\-10\-09T18:01:02\.851052\+00:00: rank \#53

2026\-10\-09T19:01:11\.250291\+00:00: rank \#53

2026\-10\-09T20:01:32\.91907\+00:00: rank \#53

2026\-10\-09T21:02:08\.376144\+00:00: rank \#55

2026\-10\-09T22:01:19\.253075\+00:00: rank \#57

2026\-10\-09T23:01:31\.990912\+00:00: rank \#57

2026\-10\-10T08:01:00\.593376\+00:00: rank \#56

2026\-10\-10T09:01:45\.527975\+00:00: rank \#58

2026\-10\-10T10:00:53\.799353\+00:00: rank \#58

2026\-10\-10T11:00:35\.65851\+00:00: rank \#57

2026\-10\-10T12:01:11\.169647\+00:00: rank \#57

2026\-10\-10T13:00:54\.008377\+00:00: rank \#55

2026\-10\-10T14:01:43\.208766\+00:00: rank \#57

2026\-10\-10T15:01:24\.825188\+00:00: rank \#57

2026\-10\-10T16:01:10\.213328\+00:00: rank \#56

2026\-10\-10T17:00:48\.84189\+00:00: rank \#55

2026\-10\-10T18:00:57\.55132\+00:00: rank \#55

2026\-10\-10T19:00:52\.379463\+00:00: rank \#54

2026\-10\-10T20:00:21\.993912\+00:00: rank \#54

2026\-10\-10T21:01:10\.657738\+00:00: rank \#54

2026\-10\-10T22:00:50\.433569\+00:00: rank \#53

2026\-10\-10T23:01:00\.95027\+00:00: rank \#54

Reflection's Beam claims frontier open\-weight coding and agentic performance with 3–4× lower inference compute, but the weights and technical report are not yet released, and the supplied discussion questions its cost,

## The brief

Reflection introduced Beam, its first open\-weight model: a sparse mixture\-of\-experts system with 501B total parameters and 23B active, aimed at coding, reasoning and agentic workloads\. It was pretrained on 23\.8T tokens and then trained with high\-compute reinforcement learning—over 100M rollouts on 10\.5K Nvidia GB300 GPUs across four weeks\. Reflection claims competitive performance with larger open models and 3–4× lower inference compute, but the weights, technical report and model card are promised later this month, and final red\-teaming and evaluations are still underway\.

- Architecture: sparse MoE with 501B total and 23B active parameters; interleaved local/global attention, fine\-grained routed experts, load balancing, and 1M\-token context after midtraining\.
- Pretraining: 23\.8T curated tokens, run in under four weeks on 6,144 Nvidia GB300 NVL72 GPUs; reported 92\.3% goodput and nine semi\-automatic rewinds\.
- RL campaign: 10\.5K GB300 GPUs for four weeks, \>100M rollouts, 256K max context, \~1\.3B sandboxes and nearly 1M coding/agentic/STEM environments\.
- Training method: fully asynchronous policy gradients with per\-token version tagging; Reflection says learning stayed stable with one\-day staleness, 107 weight versions behind\.
- Benchmark claims: competitive with GLM 5\.2 and approaching Qwen 3\.8\-Max on coding/agentic tasks; Kimi K3 remains ahead on raw capability, while Beam's claimed edge is inference efficiency\.
- Safety/alignment: a second model trained via SFT and RL, merged through multi\-teacher on\-policy distillation; reward forecasting reportedly correlated at r=0\.79 versus r=0\.46 for Best\-of\-N ceiling\.

## Discussion themes

Analyzed: 2026\-10\-06T19:00:39\.677338\+00:00

Analysis sample: Based on 15 of 25 usable stored comments. Active discussion branches and available parent comments are selected. The analysis input was further shortened to fit its context limit.

This sample may omit parts of the full thread. Selected themes do not measure community opinion or how common a view is.

### Unverifiable release and disputed generalization demo

The announcement offers no weights or technical details, prompting demands to publish weights and an HF repo and suspicion that the proprietary dataset is either problematic or unremarkable\. A demo's claim that a world\-map puzzle was only days old is challenged as the puzzle appeared earlier on LessWrong; replies further question whether search capability belongs to the model or the harness, weakening the generalization evidence\.

Sources: [Comment 49969215](<https://news.ycombinator.com/item?id=49969215>) · [Comment 49969545](<https://news.ycombinator.com/item?id=49969545>) · [Comment 49969577](<https://news.ycombinator.com/item?id=49969577>) · [Comment 49969374](<https://news.ycombinator.com/item?id=49969374>) · [Comment 49970497](<https://news.ycombinator.com/item?id=49970497>) · [Comment 49970599](<https://news.ycombinator.com/item?id=49970599>)

### Comparison with DeepSeek and other open models

Commenters compare Beam unfavorably to DeepSeek v4\.1 Flash and smaller free Chinese models, saying it is larger, more expensive, and worse on measured metrics\. One reply welcomes new entrants even if not record\-breaking, and Gemma is cited as a strong language model\.

Sources: [Comment 49969696](<https://news.ycombinator.com/item?id=49969696>) · [Comment 49969690](<https://news.ycombinator.com/item?id=49969690>) · [Comment 49969947](<https://news.ycombinator.com/item?id=49969947>) · [Comment 49970325](<https://news.ycombinator.com/item?id=49970325>)

### Size and training efficiency concerns

Beam is described as larger and more expensive to run than DeepSeek v4\.1 Flash while underperforming\. Another reply argues that using more compute and training data than predecessors without surpassing them indicates an inferior training regime that should not be encouraged\.

Sources: [Comment 49969696](<https://news.ycombinator.com/item?id=49969696>) · [Comment 49970325](<https://news.ycombinator.com/item?id=49970325>) · [Comment 49969690](<https://news.ycombinator.com/item?id=49969690>)

### Running cost and open\-weight giveaway economics

One comment notes Beam is more expensive to run than a competitor\. Others discuss the oddity and cost of open\-weight labs giving away products of millions of dollars in pre/post\-training and 10,000 GPUs running for a month under permissive licenses while facing hostile comments\.

Sources: [Comment 49969696](<https://news.ycombinator.com/item?id=49969696>) · [Comment 49972553](<https://news.ycombinator.com/item?id=49972553>) · [Comment 49975122](<https://news.ycombinator.com/item?id=49975122>)

### Proprietary training data provenance

The mention of proprietary licensed datasets draws suspicion that the data may be something the lab should not have or that it is not actually special, raising concerns about data provenance and disclosure\.

Sources: [Comment 49969545](<https://news.ycombinator.com/item?id=49969545>)

### Geopolitics and open\-weight competition

Replies frame Beam's selling point as not being from a Chinese lab, mention government contracts that forbid foreign models and an 'American DeepSeek Project,' and debate Western versus Chinese open\-weight progress, open\-source access, and broader China/West governance claims\.

Sources: [Comment 49970192](<https://news.ycombinator.com/item?id=49970192>) · [Comment 49970208](<https://news.ycombinator.com/item?id=49970208>) · [Comment 49969690](<https://news.ycombinator.com/item?id=49969690>) · [Comment 49975122](<https://news.ycombinator.com/item?id=49975122>) · [Comment 49975925](<https://news.ycombinator.com/item?id=49975925>)

## Sources & coverage

AI-generated summary · 2026\-10\-05T21:01:45\.239659\+00:00

Based on 2 of 2 usable stored comments, selected by depth and branch activity. This is a sample of the discussion. Article text may also be shortened.

Generated using deepseek\-ai/DeepSeek\-V4\.1\-Flash. Check the linked sources for full context.