Lite-Beam 1.0

A 134.5 GiB mixed 2-bit GGUF of Reflection's 501B MoE — all 256 experts retained — that runs on a 156 GB DRAM workstation with two small GPUs. With a Software+Law engram it gains +7.5 AA points on the held-out test set. The full-retrieval GPQA arm is 173/198 (87.4%), and the AIME retry arm is 29/30 (96.7%), final; its single miss hit the 32K cap. Apache-2.0.

This is a quantized derivative of reflection/Beam-1.0-501B-A23B, not a new base model. The filename carries the recognized IQ2_XXS variant; the mixed map is gate/up IQ2_XXS, down IQ2_S, attention Q6_K, router F32.

Quick links: Model file · Run locally · Mirror CLI · Engram tables · Results Space

The result

* Local held-out test 400, minimal reasoning, Beam judge. †Reflection's published FP8 reference. Methodology and provenance.

The improvement is driven by the two domains where retrieval fit:

Paired test: 38 better / 10 worse; INCORRECT falls 47 → 45.

SimpleQA-verified dev 200: base 6.5 → 50.5 with the optional wiki engram. The wiki table is a local add-on.

Benchmark comparison

The new Q2 + engram column is the full-retrieval arm. It is not a paired quant-retention comparison: the GPQA/AIME base values are short minimal screens, while the engram values are full max-effort 32K runs.

* AA and Software+Law are paired at minimal reasoning with the local Beam judge. † GPQA/AIME base screens use minimal / 512 and / 1,024 tokens; the engram rows use full max effort with a 32K budget. They are directionally useful but are not paired comparisons. GPQA retry is complete at 173/198; 194 responses are valid (173/194 = 89.2%), with 3 HTTP 500 errors and 1 length cap unresolved. AIME is final at 29/30; its single miss is a 32K length cap. HLE has 82/560 processed (28 correct; 5 HTTP errors), still running. Published BF16/FP8/NVFP4 rows are Beam's card reference on a different harness.

Full engram evaluation

The full-effort engram arms are the new comparison:

• GPQA Diamond 198 — 173/198 (87.4%) final. The retry set resolved 25/39 items; 194 responses are valid (173/194 = 89.2%), with 3 HTTP 500 errors and 1 length cap still unresolved. The base Q2 short screen is 92/198 (46.5%).
• AIME 2026 30 — 29/30 (96.7%) final at max effort / 32K. Four of the five old 8K caps were recovered correctly; the fifth hit the 32K cap. The base Q2 short screen is 9/30 (30.0%).
• HLE text-only multiple choice 560 — 82/560 processed, 28 correct, running at max effort / 32K. There are 5 HTTP errors in the current sample; raw is 34.1% and valid-answer is 28/77 = 36.4%. This is not a final score.

These are status snapshots, not final scores. The live status and methodology are on the results Space. The card will be updated when HLE finishes.

What changed

The build keeps all experts and spends bytes on attention/router correctness and the down projection. For maximum accuracy use the FP8/BF16 checkpoint; this is the 2-bit local build.

Files

SHA-256: c5d8301ffacec552338f7cfd853bc006c4fd4daa2a4e31d818b85cf686e5b47e

Run locally

Use a llama.cpp build with Beam support and the template in the GGUF. Reference layout: routed experts in DRAM; attention/dense/rest on the GPUs; ~156 GB DRAM.

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

Adjust -ncmoe and -ts for your hardware. The OpenAI-compatible endpoint is http://127.0.0.1:18360/v1. A 1M-context run was not made; serving used 131,072 tokens.

Mirror CLI

Reflection's native harness is built around Beam. It can drive this local GGUF through the OpenAI-compatible endpoint, so the native CLI uses the 2-bit weights:

uv tool install reflection-mirror

mirror \
  --api openai-chat-completions \
  --base-url http://127.0.0.1:18360/v1 \
  --model beam-1.0 \
  --context-window 131072 \
  --reasoning-effort medium

Hosted Beam through Reflection's API:

export REFLECTION_API_KEY="<your key>"
mirror --api reflection --model Beam-1.0-501B-A23B

Mirror defaults to max; for the 2-bit local build use low or medium. The first run asks you to trust the folder; accept once. See the Mirror overview and local-model setup.

Engram tables

The Software, Law, and OER tables and the retrieval code are in tablinum/beam-engram-tables. The release route is:

• Software → software + wiki
• Law → law + wiki
• Health / Finance / STEM / Humanities → closed-book

The health/StatPearls table is built locally. Its sources (NCBI Bookshelf/StatPearls, MedRAG textbooks, PubMed abstracts) do not permit redistribution of the combined passages, so the dataset includes the exact health build recipe rather than the text. Nothing from it is sent to a remote endpoint. The AA numbers are reproducible from the model plus the released Software/Law tables; the wiki root table is not included.

What you need

• ~156 GB DRAM plus the described GPU memory; 134.5 GiB of weights and KV. This is a workstation model, not a single-24-GB-GPU model.
• A llama.cpp build with Beam support and --jinja.
• 131k-token serving context.
• For maximum accuracy, use the FP8/BF16 checkpoint; this is the 2-bit local build with KLD 0.196.
• Text only. The file name carries the IQ2_XXS variant; the tensor map is mixed IQ2_XXS/IQ2_S.

Built for the Beam Build challenge

This model, the Space, and the engram tables are an entry in Reflection's Beam Build challenge (deadline November 2, 2026; $20K total across three winners).

License and attribution

Beam-1.0-501B-A23B is by Reflection, Apache-2.0. This GGUF is a modified, quantized derivative produced by Lightspeed/tablinum. Include the Apache-2.0 text and the NOTICE with any redistribution. No endorsement by Reflection is implied.

Citation

@misc{litebeam2026,
  title  = {Lite-Beam-1.0-501B-A23B-IQS\_XXS},
  author = {Lightspeed / tablinum},
  year   = {2026},
  note   = {2.30-bpw mixed 2-bit GGUF of reflection/Beam-1.0-501B-A23B},
  url    = {https://huggingface.co/tablinum/Lite-Beam-1.0-501B-A23B-IQS_XXS}
}