ProCreations/Ternary-Bonsai-2-27B-DFlash2

Bonsai 2 27B DFlash2

Updated September 18, 2026 with the best verified continued-training checkpoint: r2-dflash. The original Ternary Bonsai 2 27B PQ2_0 target is unchanged. The head is available as BF16 master weights and integer Q8_0 inference weights, not FP8.

Continued-training results

Matched testing against the previous release measured 169.12 → 170.04 decode tokens/sec (+0.54%) on eight unseen prompts with two repeats. This is a small measured gain, not a large speedup. Both arms used the same runtime and maximum 3 draft tokens.

Suite	Previous tok/s	Updated tok/s	Previous acceptance	Updated acceptance
Eight unseen prompts, two repeats	169.12	170.04	50.27%	50.71%
Six prior release prompts, two repeats	175.99	176.64	53.79%	54.20%

Acceptance is accepted draft tokens divided by proposed draft tokens. Throughput is total generated tokens divided by total decode time, including reasoning tokens. The complete comparison links to the preserved raw requests, outputs and counters in the same reports directory.

At the old fixed maximum of five draft tokens, throughput was 164.13 → 164.76 tok/s (+0.38%), with acceptance 34.86% → 35.13%. The new default of three draft tokens was selected on separate development prompts. Its higher acceptance percentage must not be attributed to training: the draft length changed. Comparing the old five-token default to the new three-token default yields 3.60% more throughput in this suite, combining weights and runtime tuning.

All tests used one RTX PRO 6000 Blackwell 96GB, one slot, context 32768, 1536 generated tokens per request, medium reasoning without a thinking-token cap, temperature 1, top-p .95, top-k 20 and min-p 0. These are short-context finite-prefix speed measurements, not broad capability scores or completed browser-task timings. Runs were sequential rather than randomized, and small gains may vary. Neither two full passes over all 8,287 training examples nor the subsequent full integer Q8 QAT pass beat the earlier winners. The three-case greedy output identity smoke passed. Previous-release objective and vision checks are historical and were not rerun on the updated weights.

Download and run

Use the supplied patched native runtime. The binary targets Linux x86-64, CUDA 13.3 and SM120 Blackwell; other hardware should build the included exact source. The head is not a standalone chat model or a standard Transformers AutoModel checkpoint.

hf download ProCreations/Ternary-Bonsai-2-27B-DFlash2 --local-dir bonsai-dflash2
cd bonsai-dflash2
sha256sum -c SHA256SUMS
python download-target.py
tar -xzf runtime/llama-bonsai-dflash2-linux-cuda13.3-sm120.tar.gz -C runtime
bash serve.sh

Bonsai-2-27B-DFlash2-Q8_0.gguf is the draft; model.safetensors and config.json are the BF16 checkpoint. The unchanged target is downloaded separately. For vision, use python download-target.py --vision and set MMPROJ to the projector path. BONSAI_MODEL can point to an existing original target.

The API binds privately to http://127.0.0.1:8080/v1. PORT, CONTEXT, DRAFT_TOKENS and LLAMA_BIN_DIR override launch settings. Medium reasoning and no thinking-token cutoff remain defaults. Build the exact native source with bash runtime/build-runtime.sh, then set LLAMA_BIN_DIR as printed by that script. Q8 quantization applies to head matrices; normalization weights retain the runtime's required floating-point representation.

Reproducibility and history

The shared experiment archive preserves source/configuration, all generated text and splits, feature metadata, selection history, full-pass coverage and raw benchmark reports for both heads. The reproduction guide explains reconstruction and the expired experiment-specific paths/deadlines. Frozen feature arrays and final optimizer/RNG states are excluded; further fine-tuning from the supplied weights starts a new optimizer. Earlier-round source is preserved alongside the current archive.

The previous release at its immutable revision retains the old weights, original donor/no-head comparisons and historical quality checks. Those earlier absolute timings must not be mixed with this matched continuation comparison. Existing historical reports remain available; their filenames do not describe the newly updated weights.

This is an independent ProCreations experiment, not an official Prism ML, Qwen or DFlash release. Original model licenses are Apache 2.0; included runtime and SpecForge sources retain their own licenses and notices. Original target: prism-ml/Ternary-Bonsai-2-27B-gguf revision 6ed5e12bf84b7a63069882c91dd9e9218647d17b. The base remains unchanged; speculative tokens are verified by it. Floating-point batching can still change text near close token decisions. See the manifest and original pinned release for donor provenance.