Anjielon/ODINO-397B-v34a-TQ1_0

ODINO 397B v3.4a

Two-plane ternary TQ1_0 · 88.2 GiB · 1.911 bits/weight
a 397B-parameter, 512-expert MoE forged on a single desktop machine

llama.cpp Vulkan TQ1_0 kernels (PR #27765) · the forge (paper + tools)

---

The headline

	ODINO v3.4a	IQ1_M (llama.cpp)
bits/weight	1.911	1.844
size	88.2 GiB	85.1 GiB
tasks solved (143 paired trials)	112 — 78.3%	79 — 55.2%

+23.1 points on real coding-and-agentic work, discordant pairs 39–6, exact McNemar p = 5.42e-07, paired-bootstrap CI₉₅ of the difference [+14.7, +31.5] points.

⚠️ The sizes are not identical and we do not round them away: the IQ1_M reference is 3.6% smaller, so the gap above comes with a small size handicap against us, not in our favour.

The finding: perplexity ranks this pair backwards

Same corpus (WikiText-2 raw, 72 chunks, ctx 4096), same engine, same machine:

	ODINO v3.4a	IQ1_M
perplexity	7.5903 ± 0.0525	7.5438 ± 0.0512
HellaSwag (same 2,108-item prefix)	75.85%	73.86%
Winogrande (1,267 items)	65.67 ± 1.33	61.09 ± 1.37
paired task suite (143 trials)	78.3%	55.2%

Perplexity — the cheapest and most widely reported of these measurements — is the only one that prefers IQ1_M. A practitioner selecting on perplexity alone would ship the file that solves roughly one task in two instead of four in five.

Task results by family

Reasoning off, two fixed seeds (7, 42), identical server and judge for both files. Frozen data: bench_v2_CONGELATO_20260829-1959.csv, md5 66c0d923c2818bb306b578a5e10cedd0.

Files

file	size	what it is
ODINO-397B-v34a.gguf	88.2 GiB	the model — two-plane ternary TQ1_0, 1.911 bits/weight
logs/*.log	—	the raw perplexity / HellaSwag / Winogrande runs behind every number above
assets/*.png	—	the figures on this page

hf download Anjielon/ODINO-397B-v34a-TQ1_0 ODINO-397B-v34a.gguf --local-dir .

Running it

The TQ1_0 Vulkan path lives in llama.cpp PR #27765 (approved, in CI). The same shaders have been exercised on AMD gfx1151 (Radeon 8060S, where this model is served daily) and on an NVIDIA RTX 4060 — no vendor-specific changes. The quantizer itself is bit-exact across CPU and CUDA backends (identical packed payloads, |Δ| = 0).

llama-server -m ODINO-397B-v34a.gguf -ngl 999 --ctx-size 4096   # ~88 GiB weights

On a 128 GB unified-memory APU the model runs fully resident. ⚠️ The TQ1_0 Vulkan kernels are not in upstream llama.cpp yet — until PR #27765 lands, build from that branch, or the model will not load.

Reproducing the numbers on this page

# perplexity (the metric that ranks this pair backwards)
llama-perplexity -m ODINO-397B-v34a.gguf -ngl 999 --ctx-size 4096 \
    --file wiki.test.raw --chunks 120

# HellaSwag / Winogrande, same settings used here
llama-perplexity -m ODINO-397B-v34a.gguf -ngl 999 --ctx-size 4096 \
    --hellaswag  --hellaswag-tasks 4000 --file hellaswag_val_full.txt
llama-perplexity -m ODINO-397B-v34a.gguf -ngl 999 --ctx-size 4096 \
    --winogrande --winogrande-tasks 1267 --file winogrande-debiased-eval.csv

The raw output of each of these runs, for this file and for the IQ1_M baseline, is in logs/.

Honesty of the measurement

• The task suite is private: aggregates are published, fixtures are not. Families built on the author's own home automation are excluded from every number on this page.
• A number without its raw log does not enter this page. The raw perplexity / HellaSwag / Winogrande logs for both files are in logs/ in this repository.
• Runs whose harness logged a transport fault (5xx, dropped connection, malformed response) are never scored as model failures: the run is discarded and re-executed.
• Four figures from an earlier night had been transcribed by hand while their logs sat in a temporary folder the system cleaned. They were re-measured from scratch rather than copied from memory — the transcribed Winogrande said 64.80; the true value is 65.67.

What is NOT yet demonstrated

The forge has one measured but not-yet-spendable lever: a Hadamard rotation over the full 256-weight TQ1_0 block that removes ~5% of reconstruction error at the same bit budget on this model's tensors (0/40 tensors worse), waiting on the decode-side kernel that undoes it. It is stated here so that the current file is not mistaken for the ceiling.

Citation

@misc{odino2026,
  title  = {Depth, Not Fidelity: What Decides Whether a Ternary Correction
            Helps a Mixture-of-Experts},
  author = {Anjielon},
  year   = {2026},
  note   = {Model: ODINO 397B v3.4a (TQ1\_0). Forge and paper:
            github.com/Anjielon/fucina}
}