DiffusionGemma B2 — experimental down-projection repair, Q4_K_M
B2 is an experimental hybrid of Edwixx DiffusionGemma and Google Base. In a frozen five-prompt, single-seed diagnostic suite, restoring 46 down-projection tensors removed the observed stray “own” substitutions and reduced the average request time from 6.275 s to 2.143 s (2.93×) versus locally quantized Edwixx on an RX 7900 XT. These are limited measured results, not a universal fix, general coding benchmark, or promise of unchanged uncensored behavior.
The original author already documents the “own” issue. Checksum-valid community weights and an independent matched local Q4 conversion both reproduced it in our setup. Re-downloading those same artifacts does not change their weights. This does not establish that every quantization, runtime, or the unquantized model behaves identically.
Download
File: diffusiongemma-26B-A4B-hybrid-B2-down-q4_k_m.gguf — 16,806,812,224 bytes / 15.652 GiB.
hf download cicero99zz/diffusiongemma-26B-A4B-B2-Experimental-Q4_K_M-GGUF \
diffusiongemma-26B-A4B-hybrid-B2-down-q4_k_m.gguf --local-dir ./b2
SHA-256: eb618d21918b1f488e8f04488c223fb2dd827ae3e497bbdb00949fcc4e932642.
Requires the experimental DiffusionGemma runtime, not an arbitrary standard llama-server. Start with runtime instructions. Verified serving is text-only, physical 32768/F16KV, public prompt+output cap 32000, no draft-MTP. Tools and images are explicitly rejected. The source architecture's broader capabilities are not advertised as tested here.
What changed
B2 was spliced at F16, then quantized with the exact 692-entry tensor-type map. It replaces 27 shared down-projections (layers 3–29) and 19 stacked expert down-projections (layers 11–29,128 experts per tensor) with Google Base payloads. The other 646 tensors retain local Edwixx payloads, including the 29 changed attention output projections. Tokenizer control-token types and the chat template are normalized to Base so the matched sailing input is the same 29 token IDs.
All 692 quantized payloads passed donor/untouched byte comparisons. The exact list, pinned revisions, hashes and metadata changes are in weights.manifest.json; reproduction instructions and the F16 splice script are included. No logit suppression, new denoising rule or retraining was used to obtain B2.
Frozen comparison
Prompt order: sailing, code, narrative, defensive security, fiction. Seed 1234, one 256-token canvas,48 maximum steps; five warmups and 15 timed requests per model. Loading, cache advisories and warmups are outside request wall time.
| Model | Exit steps in prompt order | Mean request wall time (s) |
|---|---|---|
| google-base | 10, 10, 18, 18, 16 | 2.205 |
| community-edwixx | 28, 30, 48, 48, 48 | 6.031 |
| local-edwixx | 48, 17, 48, 48, 48 | 6.275 |
| hybrid-B1-attention | 40, 20, 47, 48, 48 | 6.108 |
| hybrid-B2-down | 18, 9, 16, 12, 15 | 2.143 |
All 100 requests (25 warmups + 75 measured) matched their respective reference trajectories. Repeating the same seed measures timing consistency, not independent quality samples. Per-step timing varies with prompt/model; we do not claim kernel invariance. Fixed model order leaves a thermal/order caveat; clocks were monitored, not locked. Recorded temperature is edge, not junction.
- B2 had zero “own” assignments across the captured 256-slot trajectories in these five cases, and its generated parser passed the nested-list test.
- B1 (attention-only reversion) did not remove systemic corruption. This complementary result supports the down-projection rescue in these cases; it does not prove a unique internal causal mechanism.
- The original sailing gate was ≤16 steps; B2 took 18 and therefore failed that gate. Its clean text/latency were accepted as an operational trade-off; the original failure is retained.
- The Google Base parser also failed under this seed. The suite does not establish comprehensive programming competence for any candidate.
- Community Edwixx fell to 0.9906 GiB sampled free VRAM, below the 1.000 GiB operational floor.
REPORT.json, 75-row replicate CSV, full comparison, synthetic outputs, complete per-step canvas traces, donor checks, fixed prompts and timing records are included under evidence/. JSON derivatives remove private machine paths; their source hashes are retained in source receipts. The CSV is byte-identical to the archived original. Raw private owner/client logs and credentials are excluded.
Serving fixes and operational limits
Separate runtime patches add exact tokenization, decoded-output truncation below a canvas size, corrected completion usage, and SSE heartbeats. SSE output remains buffered. A separate ownership/drain component keeps the GPU claim until backend work ends after a client disconnect.
A subsequent 31744-token prompt plus 256-token output-canvas canary recalled all three planted markers in 18.61 s, with 1.241 GiB sampled minimum free VRAM. Browser multiturn/switch checks also passed in the local deployment; see the separate serving validation. These are bounded checks, not sustained all-context/all-client certification. Exact HTTP limits and memory configuration matter; Swift's128K/MTP/Q8 settings do not apply to B2.
Remaining limitations: only five seeded diagnostic prompts; possible residual awkward wording (one later browser response contained “de check”); no comprehensive assessment of code, refusal boundaries, languages, long-form quality or retention of Edwixx's intended uncensored behavior. Restoring Base weights can change those behaviors. The published package was checked for provenance, patch application and privacy; no new model-quality campaign was run to publish it.
Attribution and license
Independent experimental derivative by cicero99zz. AI-assisted engineering, analysis, and documentation: GPT 6 Astra. This credit does not imply OpenAI endorsement. Original models: Google DiffusionGemma and Edwixx HERETIC, at the revisions in the manifest. Their pinned model cards identify Apache 2.0; see LICENSE, NOTICE, and Google's published license. Runtime modifications retain llama.cpp's MIT license. This is not an official release or endorsement by Google, Edwixx, or llama.cpp.
The GGUF retains its historical internal general.name (Edwixx Source) so the validated bytes are unchanged. The filename, public card and manifest identify the B2 derivative; set DIFFUSION_MODEL_ID when serving it.