Lite-Beam 27B

Qwen3.8-27B in 9.26 GB (2-bit) or 6.73 GB (1-bit) GGUFs that run on small consumer GPUs. A per-model gated engram lifts knowledge far past where the quant starts, and the 1-bit build gains the most. With Tablinum-2, the 1-bit build scores +43.0 on SimpleQA (only 7 wrong of 200) and +13.5 on AA-Omniscience dev, up from −14.5. Both are the best results of the builds we tested. Apache-2.0.

This is a quantized derivative of Qwen/Qwen3.8-27B, not a new base model. The weights are built with Tablinum's calibration set, importance matrix and quantizer (the Lite-Beam method), on per-tensor type layouts credited below. Two files: IQ1_M (6.73 GB, 1-bit), the recommended build when paired with the engram, and IQ2_S (9.26 GB, 2-bit), the stronger build without retrieval. Use the v1 injection prompt (below) with either.

Quick links: Files · Run locally · Tablinum-2 engram · Lite-Beam 1.0 (501B)

The result

The index is 100 × (correct − incorrect) / n, so a wrong answer costs as much as a right one earns. The engram mostly helps by replacing confident wrong answers with grounded ones. The 1-bit build starts better calibrated (it declines instead of guessing: 71 wrong vs 149 on AA closed-book) and, with the v1 prompt, keeps that discipline when passages are injected.

These test-split arms use the Software + Law tables with the plain injection prompt on AA, and the Wikipedia table on SimpleQA. They are self-judged at low reasoning effort, so the absolute values are not comparable with published AA numbers; the paired deltas are the result. The release route is the per-domain gate with the v1 prompt (below).

Per-domain gate (AA dev 200)

Each table is kept for a model only if it helps that model in a paired dev test. v1 is the injection prompt that says: use a passage only if it states the answer to this exact question; otherwise rely on your own knowledge or say you don't know.

IQ1_M (6.73 GB):

IQ2_S (9.26 GB): closed-book −32.5 → +1.5 with the plain prompt; Health +6.1 with v1; gated ≈ +3.

The 1-bit build starts highest because it guesses less, gains about as much from the engram, and with v1 is the only build where STEM tables help. The simple route, v2 tables with v1 on every domain, scores +13.5 with no per-domain tuning. STEM and Humanities remain coverage-limited (the answer is in the top 3 passages for few questions), and better sources for both are being built for Tablinum-2.

Reasoning

10 answers still reached the 32K budget and count as wrong; on the 188 that finished, the score is 73.9%. The 1-bit build's reasoning runs were not made for this release.

What changed

KLD is measured only on assistant-turn tokens: an instruct model's system and user tokens are untrained and swamp the mean otherwise.

IQ1_M KLD was measured on the CPU; IQ2_S on the GPUs. The 1-bit build keeps 86% top-1 agreement with BF16, but its closed-book knowledge drops more, which the engram is there to recover.

IQ2_S map: attention IQ3_S / IQ3_XXS, FFN mostly IQ2_XXS–IQ3_XXS with a few IQ1 bands, output IQ4_XS, token embeddings IQ2_S, DeltaNet gates F32/BF16. The exact per-tensor maps are in the two manifests.

Files

SHA-256:

• Lite-Beam-27B-IQ2_S.gguf f08759290b3b117d3f8a7f33914148ec13b0186c6c791a5b86eaad53c362ac63
• Lite-Beam-27B-IQ1_M.gguf a2c792dc720efd4077cd1cf4c903208a924947c653b2f3188472eb979ea0ca42

Run locally

Reference rig: RTX A2000 12 GB + RTX 2060 6 GB, split 7:2.

llama-server -m Lite-Beam-27B-IQ2_S.gguf --alias lite-beam-27b \
  -ngl 99 -ts 7,2 -fa on -np 4 -c 57344 -ctk q8_0 -ctv q8_0 \
  --jinja --host 127.0.0.1 --port 18371

Faster single-stream: the EAGLE-4 draft for Qwen3.8-27B works with this quant, at +63% generation speed on our rig and close to the reference acceptance rate. It needs a llama.cpp fork with EAGLE-4 support; see quimmedes/Qwen3.8-27B-EAGLE4-GGUF. The draft is under the Eagle 4 source-available license, not this repo's, and is not bundled.

Tablinum-2 engram

The gated retrieval tables, routes and code are in tablinum/tablinum-2. Tables whose sources do not allow redistribution are shipped as exact build recipes, not text. The engram runs locally; nothing is sent to a remote endpoint.

What you need

• About 10 GB of VRAM across one or two GPUs for the 9.26 GB file with a 57K context, or about 8 GB for the 1-bit file. CPU offload works but is slow.
• A recent llama.cpp with Qwen3.5/3.8 (qwen35) support and --jinja.
• Optional: the Tablinum-2 tables (sizes vary per domain) and its retrieval code.
• For maximum accuracy use the BF16 / FP8 checkpoint. This is the 2-bit local build.

License and attribution

Qwen3.8-27B is by the Qwen team (Alibaba Cloud), Apache-2.0. This GGUF is a modified, quantized derivative produced by Lightspeed / Tablinum. The per-tensor type layout follows ISTA-DASLab/Qwen3.8-27B-GSQ-RCO-GGUF (Apache-2.0); the tensor data is our own quantization. The IQ1_M build follows the per-tensor type layout of unsloth/Qwen3.8-27B-GGUF UD-IQ1_M (Apache-2.0); again the tensor data is our own. Include the Apache-2.0 text and the NOTICE with any redistribution. No endorsement by Qwen, ISTA-DASLab or Unsloth is implied.

Citation

@misc{litebeam27b2026,
  title  = {Lite-Beam 27B},
  author = {Lightspeed / tablinum},
  year   = {2026},
  note   = {Mixed 2-bit GGUF of Qwen/Qwen3.8-27B with a per-model gated engram},
  url    = {https://huggingface.co/tablinum/Lite-Beam-27B}
}