Beam,
on your desk.
A 134.5 GiB mixed 2-bit GGUF of Reflection's 501B MoE. All 256 experts retained. It runs on a 156 GB DRAM workstation — and the engram takes it +7.5 AA points, to 173/198 GPQA and 29/30 AIME.
The engram effect.
Same model, same test. The only change is a local Software+Law retrieval route.
| arm | correct | incorrect | index |
|---|---|---|---|
| Base Q2 | 112 | 47 | 16.25 |
| Base + S+L engram | 140 | 45 | 23.75 |
400 public test questions · minimal reasoning · Beam judge. *with optional Software+Law engram.
| benchmark | Q2 base screen | Q2 + engram | FP8 published |
|---|---|---|---|
| AIME 2026 | 9/30 (30.0%) | 29/30 (96.7%)† | 98.0% |
| GPQA Diamond | 92/198 (46.5%) | 173/198 (87.4%)† | 90.3% |
| AA-Omniscience | 16.25 | 23.75* | 8.8 |
| HLE (no tools) | not published | 82/560 (28 correct), running | 36.0% |
| LiveCodeBench | not run | not run | 88.1% |
| IFBench | not run | not run | 80.7% |
† AIME/GPQA base rows are short minimal screens (1,024 / 512 caps); engram rows are full max-effort 32K runs, not paired. GPQA retry is complete at 173/198 (194 valid; 3 errors, 1 length cap). AIME is final at 29/30; its single miss hit the 32K cap. HLE has 82/560 processed (28 correct; 5 HTTP errors), running. AA and SimpleQA are paired. FP8 column is Beam's published card reference.
The full arm has a column now.
Same Q2 artifact, full retrieval, max effort. The short base screens are shown separately so the effort mix is explicit.
Status snapshot at last update. GPQA/AIME engram rows are full max-effort 32K; their base rows are short minimal screens, so those deltas are mixed-effort, not paired. HLE exactMatch is a separate judge run.
Retrieval that knows its place.
Software and Law get the tables. Everything else stays closed-book. That is the route that held on the untouched test.
Health and OER were tested and did not add on this prompt. The health table is built locally; its source terms do not permit redistribution, so the dataset includes the exact build recipe.
| table | passages | status |
|---|---|---|
| software | 4,827,247 | released |
| law | 15,354,819 | released |
| OER | 17,500,842 | released |
| health | 27,390,442 | build recipe |
One GGUF. One command.
Routed experts in DRAM, attention and dense tensors on the GPUs. Built for a 156 GB workstation.
llama-server \ -m Lite-Beam-1.0-501B-A23B-IQ2_XXS.gguf \ --alias beam-1.0 \ -ngl 99 -ncmoe 99 --load-mode none \ -t 38 -tb 38 -c 131072 -b 4096 -ub 4096 \ -ctk q8_0 -ctv q8_0 -np 2 --kv-unified -fa on --jinja \ -ts 3,1 --cache-ram 8192 \ --host 0.0.0.0 --port 18360 --no-webui
OpenAI-compatible endpoint: 127.0.0.1:18360/v1
Apache-2.0 · quantized derivative of reflection/Beam-1.0-501B-A23B. SHA-256 c5d8301f…686e5b47e.
Mirror CLI.
Reflection's terminal coding agent, pointed at the local Q2 endpoint. The native way to drive Beam.
Built for the Beam Build challenge.
Model + Space + engram tables. Deadline Monday, November 2, 2026 · $20K across three winners.