CyberdyneLabs Program 01 / 06
284 B parameters · 13 B active · 1 M token context · MIT license. Frontier reasoning on one consumer card. We open-source the inference reference that produces it correctly, the architectural findings that make it work, and the runtime layer that makes the same hardware do more.
Below: the Python reference (verified · Paris top-1 with +11.13 logit margin), the 8 architectural gotchas that broke our own ports, and downloadable artefacts.
Achievements
Get the tools
Full downloads index → /downloads · All 66 rendered reports → /r/ · Full history (V4-Flash → ADAM × ARC-AGI-3) → /history
About
The world has produced extraordinary open foundation models, and almost no rigorous practice for shaping them after release. Our laboratory exists to close that gap.
Our active patients are open foundation models — Qwen 2.5 0.5B as the lower-organ donor and Qwen 2.5 7B as the top-brain — alongside an autopsy archive of larger artefacts including DeepSeek V4-Flash. The standard answer to making any such model behave the way you need it to is fine-tuning — and fine-tuning, in practice, is hope dressed in machinery. You feed in data; you wait; you ship whatever comes out; you cannot say with certainty what you changed.
Surgery operates differently. We treat the weight space as anatomy. We open the model up, we identify the structures responsible for a behaviour we want to change, and we modify them — locally, measurably, reversibly. Where the field at large grows models by accumulation, we refine them by intervention.
The work spans every layer of the stack. The compression formats that let very large models run on small machines. The native compiled runtime that replaces the research-grade scripting most laboratories ship with. The internal memory that gives a refined model something resembling continuity. The verifier that tells us, on every output, whether the answer is supported. None of this exists in isolation. All of it serves the same goal: to produce systems whose behaviour we understand and whose claims we can defend.
The Stack
At the centre of every system we ship sits a single high-capacity reasoning model, derived from an open foundation through targeted intervention. It carries the breadth of a frontier model and the discipline of a system whose behaviour was shaped one circuit at a time.
Around the central model sits a fleet of compact specialists — each one a sub-billion-parameter expert at a single narrow function: structured output, code skeleton, claim extraction, contradiction analysis. They run cheaply, in parallel, and answer to the model above them.
A reasoning system without memory begins every encounter from zero. Ours does not. A structured archive of hundreds of indexed volumes, addressable down to the line, gives the assembled system something other models lack: a continuous record of what it has thought, what it has been told, and where each fact came from.
Every request flows through a self-organising routing layer that decides which organs to wake, which memories to retrieve, and which paths to reinforce or starve. It is, in effect, a circulatory system — quiet, adaptive, and the reason the assembled body responds as one.
No claim leaves the system without passing a strict verifier. Every assertion that references memory must carry a pointer to its source. Every output that cannot be backed up is flagged as such, in plain language, before it reaches the user. The default, in our laboratory, is suspicion.
Most laboratories ship Python. We ship a compiled native runtime. Memory, model loading, attention kernels, tiering between fast and slow storage — all of it written in low-level systems code. The result is a complete model deployed on a single consumer GPU: less hardware, less latency, no scripting language between the model and the machine.
Selected Work
In the V4-Flash autopsy line (Q1 2026), we took DeepSeek V4-Flash — an open-weight Mixture-of-Experts foundation model with 284 billion parameters in total, 13 billion active per token, a one-million-token context window, an MIT license, and an inference profile that ordinarily requires multiple data-centre-class GPUs to operate — and we drove it through end-to-end inference on a single consumer GPU.
Measured: 1.86 tok/s decode, 2.08 sec/tok on the 18-layer text decode after expert-streaming repack, with 89 % of wall time in expert I/O. The artefact taught us that the bottleneck is not arithmetic — it is the choreography of moving specialised experts in and out of working memory. Currently active: our working Python inference reference now produces the correct Paris top-1 answer with a +11.13 logit margin on the same RTX 3060 Ti — released as open-source pipeline with 8 documented architectural findings.
Targeted excision on Qwen 2.5 0.5B (Physarum-05B donor tissue): 14.94 % global / ≈ 22 % per-tensor weights zeroed, guided by an internal signal that identifies parameters with no measurable contribution. The behaviour delta was not invisible. Perplexity rose +12.5 %; MMLU-mini lost −22 %; GSM8K-mini lost −20 %. JSON-schema and code-skeleton smoke survived. Throughput was preserved.
This was the proof of principle, not the proof of victory. A model is not a single inseparable thing — but it is also not a free dinner. Healthy and dead tissue can be distinguished; what survives the operation depends on the gate you used. We publish the deltas instead of hiding them, because the trajectory matters more than any one pass.
The single largest cost in operating a model the size of DeepSeek V4-Flash on consumer hardware is not arithmetic. It is the choreography of moving the model's hundreds of specialised experts in and out of working memory, again and again, on every forward pass.
By reorganising the way these specialised parts are packed and streamed from disk, we turned the most expensive operation in the inference pipeline into a tractable one. The same model, on the same hardware, ran roughly six times faster on its decode loop. The format is internal, instrumented, and reproducible. We use it now as the substrate for everything else.
The current memory spine is a structured archive of 305 source files / 58 996 lines: organ-surgery transcripts, docs, reports. Every line carries a sha256[:16] address; the manifest lives at data/memory_spine/manifest_v1.jsonl. Breakdown: 45 organ-surgery files (10 863 lines), 20 docs (4 783 lines), 240 reports (43 350 lines).
Indexing is shipped. Exact-lookup CLI and TF-IDF semantic ranker are queued, not done — we say so explicitly. It is not a chat history and it is not a vector store. It is the spine on which exact-recall and contradiction-detection will sit.
On the Table
Each patient that enters the runtime carries four artefacts: a baseline number, a post-surgery delta, a verifier-checked anchor set, and a frozen pack hash. No merge without those four.
| Patient | Size | Role | Status |
|---|---|---|---|
| Qwen 2.5 0.5B (Physarum-05B) | 0.5 B / 988 MB · Apache 2.0 | lower-organ base | in production · multiple specialised packs |
| Qwen 2.5 7B (Physarium-7B Q4) | 7 B / 5.55 GB Q4 · Apache 2.0 | top-brain · 7B fallback | in production · 83.58 tok/s (llama.cpp) · 18.27 tok/s native default · RTX 3060 Ti |
| DeepSeek V4-Flash | 284 B total / 13 B active · MoE · 1 M context · MIT | frontier MoE host · expert-streaming · native inference engine | Python reference active · Paris top-1 +11.13 logit margin · native engine in active development |
| DeepSeek V4-Pro | 1.6 T total / 49 B active · MoE · 1 M context · MIT | frontier reference · agentic coding (SWE-Bench frontier tie) | instrumented for runtime study · architectural diff vs V4-Flash captured |
| Qwen 3.5 small (0.6B · 1.7B · 4B) | 0.6–4 B · Apache 2.0 | next-gen organ base · replaces Qwen 2.5 0.5B donor line | queued · BD10 sweep |
| Qwen 3.5 mid (8B · 14B · 32B) | 8–32 B · Apache 2.0 | top-brain candidate · 7B fallback successor · 32B = best small dense coder | queued · BD11 · scoping vs Physarium-7B |
| Qwen 3.6-27B-Coder | 27 B dense · Apache 2.0 | code-organ specialist · best small dense coder under Apache-2.0 | queued · code-organ BD candidate |
| Qwen 3-235B-A22B | 235 B total / 22 B active · MoE · Apache 2.0 | flagship MoE organ host · safest enterprise license · expert-streaming target | queued · MoE BD12 candidate |
| Llama 4 Scout | 109 B total / 17 B active · MoE · 10 M context · Llama license | long-context multi-organ scaffold candidate | scoping · license review pending |
| Llama 4 Maverick | 400 B total / 17 B active · MoE · Llama license | frontier-tier MoE candidate · larger experts vs Scout | scoping · expert-streaming bench candidate |
| Gemma 4 E2B / E4B | 2 B / 4 B effective · Apache 2.0 | edge / mobile organ class | queued · multimodal scoping |
| Gemma 4 26B MoE | 26 B / 3.8 B active · MoE · Apache 2.0 | MoE-organ research host · pair with V4-Flash native work | queued · expert-streaming bench candidate |
| Gemma 4 31B Dense | 31 B · Apache 2.0 | top-brain candidate · 70B-class behaviour at consumer-VRAM cost | queued · vs Qwen 3.5 32B head-to-head |
| Phi-4 Reasoning Plus | 14 B · MIT | reasoning organ · rivals far larger models on complex reasoning | queued · reasoning-organ BD candidate |
| Phi-4 Mini | 3.8 B · 128 K context · MIT | edge organ · long-context on resource-constrained hardware | queued · edge-organ candidate |
| Mistral Large 3 | flagship dense · Apache 2.0 | top-brain candidate · open re-licensing of Mistral flagship | scoping |
| Mistral Small 4 | 24 B · Apache 2.0 | efficient dense organ · code + general | queued · vs Qwen 3.5 32B |
| OLMo 2 | 7 B / 13 B · fully open (data + checkpoints + logs) · Apache 2.0 | research-grade donor · only family with reproducible training trace | scoping · ideal for surgery instrumentation |
| Kimi K2.6 (Moonshot) | open-weights | frontier candidate · #1 Artificial Analysis Index (54) · long-context | scoping · license review |
| GLM-5.1 (Z.ai / Zhipu) | open-weights · MIT | multilingual organ · cleanest MIT license among Chinese frontier opens | scoping |
| Nous Hermes 4 | Llama 4 base · community fine-tune | chat / instruction / persona-tuned organ candidate | scoping |
| DeepSeek-R1-Distill-Qwen-1.5B | 1.5 B · MIT | reasoning-organ candidate | scoping |
Native Speed Ladder
Every step measured on the same RTX 3060 Ti, Physarium-7B Q4. Each row is a milestone in the native runtime; each number has a report file. The ladder is the headline arc of Phase 6 → Phase 13.
| Phase / Configuration | Speed | vs prev | Note |
|---|---|---|---|
| V4-Flash 284B PyTorch warm decode | p50 9.6 s/tok | — | flagship demo · 8 GB VRAM |
| Physarum-05B-Organic baseline | 27.15 tok/s | — | 0.5B BF16 baseline |
| CPU baseline · 0.5B | 1.91 tok/s | — | reference floor |
| CUDA full GPU 0.5B (Phase 8E.1) | 116 tok/s | 61× CPU | byte-identical |
| CUDA fused 7B BF16 streaming (8E.2) | 0.20 tok/s | — | correctness proof, not main path |
| Q4 NUCLEAR resident 7B (8E2) | 11.16 tok/s | 280× CPU baseline | 5.55 GB Q4 group=128 · 28 layers in VRAM |
Q4 native v2 default --chat | 18.27 tok/s | +64 % vs NUCLEAR | — |
| Q4 native + DP4A=1 (opt-in) | 28.99 tok/s | +59 % | — |
| Q4 native + DP4A · tg128 | 41.69 tok/s | +44 % | — |
| llama.cpp backend (LLAMACPP_URL) | 83.58 tok/s | +100 % | production · clean-room autopsy |
| Mode C llama.cpp acceptance · mean wall | 2.99 s | — | per query, 18-task suite |
sources: reports/EXTERNAL_BACKEND_SHOOTOUT_V2.md · reports/PHASE_8E8A_DP4A_NATIVE_BACKEND.md · reports/CURRENT_TRUTH_LEDGER.md §2 · 5-run mean on RTX 3060 Ti
Surgery Cycle Ledger
Eight surgery passes were reverted before the production code-skeleton organ was kept. We publish the trajectory because failed passes are the proof that the gate doctrine is real, not a slogan.
| Pass | Lever / variant | Outcome |
|---|---|---|
| BD6 pass-1 | physarum05b_code_skeleton.planck | KEPT · production · 13/100 MBPP, 6/164 HE, anchor 19/19 |
| BD6.2 | retraining | REVERTED · overtrain, MBPP regressed 13 → 6 |
| BD6.3 | anchor gate | REVERTED · failed gate |
| BD6.4 | anchor positive partial | REVERTED · partial |
| BD6.5 | stratified poison 15/19 | REVERTED · 13/19 anchor |
| BD6.6 | stratified mix v6 | REVERTED · over-anchor regression |
| BD6.7 | KL-anchor ladder λ=0.10/0.20 | REVERTED · no lift |
| BD6.8D | token-weighted CE | REVERTED · no lift |
| BD6.8D2 | per-bench poison + asymmetric holdout | REVERTED · over-tuned |
| BD6.8D-rank | r=16 / α=32 | FREEZE DECISION · ship pass-1, freeze BD6.x |
| BD7 | triz_contradiction_v2.planck | KEPT · 88/100 strict 6-field JSON, fallback 0 |
| BD8 V1–V5 | critic_lite + wound (ARIZ rescue path) | BLOCKED · rescue 0/n on ARIZ JSON · wound v2 retained for in-chat rescue |
| BD9 | phys05_json_repair | KEPT · 10 / 10 GREEN on production failure catalog · loss 0.055 → 0.0003 · 280 rows |
| BD9 | phys05_claim_extractor | KEPT · GREEN · clean structured JSON · loss 0.51 → 0.04 · 25 hand-curated rows |
| BD9 | phys05_test_writer | YELLOW · pytest shape correct, semantics drift (currying confusion + Human-token leak) |
| BD9 | phys05_cache_matcher | YELLOW · correct integer + post-answer drift · runtime regex extracts head |
| BD9 | phys05_renderer | RED · output corrupted · loss did not converge (0.69 ceiling on 25 rows) · queued BD9.1 |
Open-source reference
We are releasing the full Python reference pipeline for DeepSeek-V4-Flash inference — the same code that produces the correct Paris top-1 answer with a +11.13 logit margin on a single RTX 3060 Ti via WSL2.
This is the reference oracle: anyone porting V4-Flash to a different runtime (CUDA, Triton, MLX, Rust, anything) can use it to cross-check activations layer-by-layer with compare_dumps.py. Architecture nuances that broke our own ports are documented openly in PYTHON_PIPELINE_DOC.md so nobody has to rediscover them.
License: MIT · Model weights: DeepSeek-AI MIT · Verified 2026-05-31
Download
flash_mvp.py · flash_mvp_chat.py · kernel_pytorch.py · dump_ref_v4.py · compare_dumps.py · fht_fallback.py · doc8 architecture findings that broke our own ports
Reading model.py isn't enough — the model gives garbage without each of these. Documented openly so nobody has to rediscover them.
| # | Subsystem | What's easy to get wrong |
|---|---|---|
| 1 | Chat template | Without the trailing <｜Assistant｜></think> wrapper, the model emits garbage. Non-thinking mode is encoded by that closing </think>. |
| 2 | Hash routing computes original_scores | Layers 0/1/2 route via tid2eid, but weights still come from sqrt(softplus(x · Wᵀ)) followed by gather + normalisation + scaling factor 1.5. Naive uniform 1/top_k weights destroy magnitudes. |
| 3 | act_quant double-apply trap | For GEMM input, act_quant returns (y_raw, scale) and the GEMM applies scale internally. For inplace simulation, it writes back y · scale. Confusing them = silent 1000× underflow in MoE. |
| 4 | Compressor overlap a/b split | CSA (m=4) compressor weights split wkv, wgate, ape into "a-portion" (previous chunk in overlap window) and "b-portion" (current chunk). HCA (m=128) has no overlap → single weight. |
| 5 | mHC C_l = 2 · sigmoid | Output mapping for Manifold-Constrained Hyper-Connections is 2 × sigmoid (range [0, 2]) — the factor 2 is critical and easy to miss. |
| 6 | SwiGLU asymmetric clamp | Up-component clamped to [-10, +10]. Gate component capped at +10 only — no lower cap. The asymmetry is real and intended. |
| 7 | Inverse RoPE on attention output | After attention, before the grouped output projection, apply inverse RoPE to the last 64 dims of the attention output ("Partial RoPE", paper §2.3.3). Missing this destroys long-context coherence. |
| 8 | Per-head Q RMS without learned weight | After wq_a → q_norm → wq_b, apply per-head RMS with no learned weight: q *= rsqrt(q.square().mean(-1) + eps). Skipping it silently degrades head specialisation. |
Full algebraic derivations + per-layer activation reference (21 layer-0 dumps + 9 deeper layer inputs + final logits + pre-head) in PYTHON_PIPELINE_DOC.md §4 and §5.
Reproduce Paris in 4 commands
tar xzf python_v4_paris_pipeline.tar.gz && cd v4_pipeline
pip install torch transformers safetensors numpy
export V4_MODEL=/path/to/v4_original # HF checkpoint with safetensors + inference/
python3 flash_mvp_chat.py --ckpt $V4_MODEL \
--user "What is the capital of France? Answer in one word." \
--mode chat --max-seq 64
# → predicted token 51119 ('Paris'), logit +40.75
Generate reference dumps for porting validation: REF_N_LAYERS=43 python3 dump_ref_v4.py (~11 min on RTX 3060 Ti).
Reports
See all 66 rendered reports → /r/ · Full 95-report archive → /downloads
Open Tools
The compiled inference loop. mmaps .planck packs, runs CUDA forward, orchestrates organs. Acceptance suite 18/18 on llama.cpp backend, identity probe 14/14, integrity audit 10/10. Q4 7B at 5.55 GB VRAM, 83.58 tok/s production (llama.cpp) · 18.27 tok/s native default · 28.99 tok/s with DP4A flag · RTX 3060 Ti.
Failure→repair pair forge, 7B-teacher → student data builders, QLoRA training drivers, adapter merge + planck repack. Every tool produces .planck / .jsonl / .md artefacts; runtime never imports them.
Indexing is shipped: every line of the spine has a sha256[:16] address. Exact-lookup CLI and TF-IDF semantic ranker are in build, not done. We say so explicitly.
Hard verifier in place across all production routes. Source-pointer-required gate runs on memory-anchored seeds (14/14 stretch passed); not yet enforced on free-form chat replies. Honest scope.
Principles
Where the field grows models by accumulation, we refine them by intervention. Local, targeted, measurable. We change the smallest set of parameters that produces the change we want, and we know which ones.
Research code belongs in the laboratory. Production systems belong in compiled native code. The translation is not optional and it is not a future concern. It is the work.
No claim leaves the system without a pointer to evidence. No memory is trusted without provenance. No operation is shipped without a reproducible benchmark. The default in this laboratory is doubt.
Honest Current Status
gigachad_native · single binary · single GPU.