Vita0818/Vireqo-27B-Plus-260816

Vireqo-27B-Plus-260816

Unofficial, experimental, text-only, Ternary-Bonsai-guided cross-version GGUF derived from both Qwen3.8-27B and Ternary-Bonsai-27B.

Vireqo-27B-Plus-260816 is a 7.06 GiB ternary 27B-class GGUF built for constrained Apple Silicon systems. It preserves the trained {-1, 0, +1} code pattern of Ternary-Bonsai-27B, conservatively refits block scales toward the aligned Qwen3.8-27B weights, and retains source-consistent recurrent controls and token interfaces. The release was built on a Mac with approximately 12 GiB reserved for the task and was loaded successfully by LM Studio’s bundled llama.cpp backend.

This is not a conventional direct quantization of Qwen3.8-27B, not an official Qwen or Prism ML release, and not a claim that Qwen3.8 quality is retained. The large low-bit matrices are primarily the trained Ternary-Bonsai prior. Qwen3.8 contributes a limited scale-fitting term, selected F32 control tensors, and release tokenizer/template information. The complete method and its limitations are disclosed below and in `TECHNICAL_README.md`.

中文摘要：这是 Vireqo-27B-260816 的三值 Plus 分支。它用更大的约 7.06 GiB 文件保留 Ternary-Bonsai 的 -1/0/+1 离散权重结构，并让主干 scale 以 25% 权重向 Qwen3.8 拟合。正式文件在 LM Studio 同款后端中通过了短答、基础长回复、384-token 压力测试与 768-token 扩展测试的复读稳定性验收；但严格计数、格式、事实正确性与 Thinking 模式均没有被接受为可靠能力。

Release file

Exact construction

The release uses one fixed layout:

For each 64-weight block in the 496 fitted backbone tensors, let z = code - 1, so z ∈ {-1, 0, +1}. The release computes:

d_fit   = max(dot(W_qwen, z) / dot(z, z), 0)
d_final = 0.75 * d_ternary_bonsai + 0.25 * d_fit
W_plus  = d_final * z

If a block contains only zero codes, it retains the source scale. The packed Ternary-Bonsai code bytes are copied unchanged for every Q2 tensor:

• —Q2 code migrations: 0;
• —illegal fourth code (code=3): 0;
• —Q2 blocks scanned: 420,208,640;
• —token_embd.weight: byte-exact Ternary-Bonsai Q2 payload;
• —output.weight: byte-exact Ternary-Bonsai Q2 payload;
• —192 .ssm_a, .ssm_conv1d.weight, .ssm_dt.bias, and .ssm_norm.weight tensors: byte-exact Ternary-Bonsai F32 payloads;
• —remaining 161 F32 normalization/control tensors: converted Qwen3.8 values;
• —tokenizer tokens and BPE merges: verified aligned;
• —chat template: pinned Qwen3.8 download;
• —vision, mmproj, MTP, and external draft weights: excluded.

The 25% Qwen fitting weight is a conservative engineering prior inherited from the stable Q1 route. It is not claimed to be globally optimal. A sampled reconstruction audit measured Qwen-target NMSE changing from 0.44395375 for the original Ternary-Bonsai code+scale to 0.43573998 for the fixed-code blended scale, a relative reduction of 1.8501%. This is a weight-space diagnostic, not a 1.85% benchmark or quality claim.

Why “Plus”

“Plus” identifies the ternary size/representation branch. It does not mean the Qwen mix is more aggressive, and no standard benchmark comparison against the Q1 release is claimed here.

Embedded generation defaults

The final GGUF contains:

The first three values follow the upstream Ternary-Bonsai recommendation. The modest repetition penalty and 64-token window are release safeguards for sustained low-bit generation. A final llama.cpp regression omitted all sampling fields and confirmed that these metadata defaults produced 8/8 nonempty base responses without detected loops. A frontend’s saved per-model preset may override the GGUF values.

Evaluation summary

The tests used LM Studio’s installed llama.cpp-mac-arm64-apple-metal-advsimd-2.28.2 backend, full GPU offload, flash attention, Q8_0 K/V cache, one parallel request, and reasoning disabled.

The detector is intentionally conservative. A flag can be raised when a required term appears four times even if the answer is progressing normally. Manual review found no repeated sentence, duplicated paragraph, or tail cycle in the 384- and 768-token suites.

Representative results:

• —short factual probe: 法国的首都是巴黎。, 7 completion tokens, normal stop;
• —final-default base suite: 8/8 nonempty, zero detected loops, zero thinking-tag leakage;
• —final-default difficult subset: lighthouse story, generative-AI essay, and student/teacher dialogue all stopped normally;
• —one 768-token Chinese report reached its token cap with repeated-8gram fraction 0.000;
• —one 768-token English tutorial reached its cap while continuing through new sections; repeated phrases came from the required running bookstore example.

Stability is not quality equivalence. Observed weaknesses include mixed-language fragments, invented numerical claims, incomplete strict counts, and generation-limit truncation. In the final student/teacher test, the model stopped normally but did not produce exactly twelve requested turns.

Local runtime result

These values are one-machine evidence. LM Studio versions, context size, concurrent applications, unified-memory pressure, and macOS caching can change the result.

The bundled backend can terminate with status 1 or 134 after receiving a duplicate interrupt during server shutdown. All cited model loads, responses, and timing records completed before cleanup; this shutdown behavior was not classified as an inference failure.

Recommended LM Studio settings

1. 1.Load Vireqo-27B-Plus-260816.gguf with a current llama.cpp runtime.
2. 2.Start with context 2048; reduce to 1024 if the system is under memory pressure.
3. 3.Use maximum GPU offload, flash attention on, Q8_0 K and V cache, and parallel requests 1.
4. 4.Keep Thinking / Reasoning off. This release did not accept thinking mode.
5. 5.Do not attach an mmproj, MTP model, or speculative drafter.
6. 6.Let the GGUF defaults apply initially: temperature 0.7, top-k 20, top-p 0.95, repeat penalty 1.08, repeat window 64.
7. 7.Start with a 256-token maximum output. Increase toward 384 or 768 only when the task needs it; reaching a cap is truncation, not a loop.

First diagnostic prompt:

请只用一句完整的中文回答：法国的首都是哪里？

Expected behavior is a short answer equivalent to 法国的首都是巴黎。 followed by a normal stop.

See `LM-STUDIO-使用指南.md` for the no-copy local installation, GUI settings, and troubleshooting.

llama.cpp example

GGML_METAL_TENSOR_DISABLE=1 ./llama-cli \
  -m Vireqo-27B-Plus-260816.gguf \
  -ngl all -c 2048 -fa on -ctk q8_0 -ctv q8_0 \
  --jinja --reasoning off \
  --temp 0.7 --top-k 20 --top-p 0.95 --min-p 0 \
  --repeat-penalty 1.08 --repeat-last-n 64 \
  -n 256 -st \
  -p '请用一句完整的中文回答：法国的首都是哪里？'

GGML_METAL_TENSOR_DISABLE=1 was required by the separately pinned local build on the tested M5 setup. LM Studio’s bundled backend handled the unavailable Metal Tensor API path automatically. This is a runtime-specific detail, not evidence of GGUF corruption.

Limitations and non-claims

• —Thinking/reasoning is not accepted for this release; keep it disabled.
• —Strict sentence, paragraph, turn, and word counts are unreliable.
• —Some long responses stop early or reach their configured token cap.
• —Grammar errors, mixed-language fragments, omissions, hallucinated numbers, and factual errors remain possible.
• —Context quality beyond 2048 was not accepted. Architecture metadata is not a usability guarantee.
• —Knowledge retention, reasoning, math, coding, tool use, safety, and standard benchmarks were not comprehensively evaluated.
• —This artifact is not equivalent to Qwen3.8-27B BF16, a conventional Q4/Q5 quantization, or the original Ternary-Bonsai model.
• —Do not use it for medical, legal, financial, security-critical, or other high-stakes decisions.

The artifact should be treated as a reproducible cross-version ternary-prior-transfer experiment with encouraging long-output stability, not as a dependable general-purpose 27B assistant.

Source revisions

The exact Ternary-Bonsai input was Ternary-Bonsai-27B-Q2_g64.gguf, 7,585,330,240 bytes, SHA-256 59a45d1ecef702b14531b06d22949f33b25c1897da31a8c0b298e01e4d9138eb.

Reproducibility files

• —Vireqo-27B-Plus-260816.gguf.provenance.json — final release provenance;
• —weight-build.provenance.json — complete weight-build and per-tensor statistics;
• —metadata-finalization.provenance.json — proof that release metadata rebuilding preserved every tensor payload byte;
• —artifact-validation.md / .json — independent structural and source validation;
• —runtime-validation.json — runtime, memory, suite, and acceptance summary;
• —long-regression-*.md / .json — raw base, stress, extended, and final-default response records;
• —source-revisions.json — pinned sources;
• —build-info.txt — compact build handoff;
• —TECHNICAL_README.md — full technical route and recovery ladder;
• —LM-STUDIO-使用指南.md — local no-copy installation and troubleshooting;
• —SHA256SUMS.txt — release checksum.

License, attribution, and naming

The artifact is provided under Apache License 2.0, consistent with the pinned upstream repositories inspected for this build. Review LICENSE and NOTICE before redistribution and re-check upstream terms if using different revisions.

• —Qwen3.8-27B: Copyright 2026 Alibaba Cloud, Apache License 2.0.
• —Ternary-Bonsai-27B: Copyright 2026-present Prism ML, Inc., Apache License 2.0.
• —Retained Bonsai attribution: Created using Bonsai by Prism ML.

Vireqo-27B-Plus-260816 is the name of this independent experimental artifact. It is not affiliated with, sponsored by, or endorsed by Qwen, Alibaba Cloud, Prism ML, or Bonsai. Those names are used only to identify the disclosed upstream sources and lineage.