Training-free, closed-form 9-Subspace Weight-Space Programming (WSP v3.0) for Qwen/Qwen3.5-4B (32 layers, hidden_size=2560, intermediate_size=9216, swiglu MLP, 1+w RMSNorm, synthesis_mode=hybrid_bf16_forward_hooks_s1_to_s6_plus_proxy_s7_to_s9), evaluated end-to-end across 100 unified benchmark tasks (30 Disjoint Holdout across S1..S9 + 70 Established Public Benchmarks) and 20 live stateful Terminal-Bench 2.0 container tasks in WSL2.
[!IMPORTANT] Uniform 100-Task + Terminal-Bench 2.0 Result (
qwen35_4b):
- Unified 100-Task Suite: Base
72/100 (72.0%)-> WSP v3.089/100 (89.0%)(+17.0 pp)- Holdout (
S1..S9,30tasks):24/30 -> 29/30- Established Public (
70tasks:20GSM8K,15HumanEval,15BFCL Irrelevance,20NL2Bash):48/70 -> 60/70- Terminal-Bench 2.0 (
20live WSL2 bash tasks): Base3/20 (15.0%)-> WSP4/20 (20.0%)(+5.0 pp;107.7 tok/s)- Speed & VRAM (
-c 16384 -ctk q8_0 -ctv q8_0):117.0 tok/s, load delta3269 MiB, peak delta3370 MiB- Cross-Hardware Fit: RTX 2060 (
6 GB): Yes (~2.7 GiB spare @ 16k) | RTX 4070 Ti (12 GB): Yes (~8.0 GiB spare @ 16k)
Base -> WSP v3.0)| Suite | Tasks | Base Pass | WSP v3.0 Pass | Avg Completion Tokens (Base → WSP) |
|---|---|---|---|---|
S1_THINK (Holdout S1: Think Closure) |
3 |
0/3 (0.0%) |
3/3 (100.0%) |
1024.0 → 695.7 |
S2_RESTRAINT (Holdout S2: Tool Restraint & Anti-Sycophancy) |
4 |
4/4 (100.0%) |
4/4 (100.0%) |
404.2 → 424.2 |
S3_ROUTER (Holdout S3: Multi-Tool Router) |
5 |
5/5 (100.0%) |
5/5 (100.0%) |
137.6 → 133.8 |
S4_CODE (Holdout S4: Mid-Band Code Synthesis) |
2 |
2/2 (100.0%) |
2/2 (100.0%) |
884.0 → 520.5 |
S5_TURN2_JSON (Holdout S5: Turn-2 Structured JSON) |
3 |
3/3 (100.0%) |
3/3 (100.0%) |
127.3 → 33.3 |
S6_HANDOFF (Holdout S6: Escalation Handoff Gate) |
4 |
4/4 (100.0%) |
4/4 (100.0%) |
178.2 → 297.2 |
S7_INJ_GUARD (Holdout S7: Indirect Injection Guard) |
3 |
2/3 (66.7%) |
3/3 (100.0%) |
148.0 → 225.0 |
S8_GROUNDING (Holdout S8: Grounded Abstention) |
3 |
1/3 (33.3%) |
2/3 (66.7%) |
113.7 → 88.3 |
S9_ERR_RECOV (Holdout S9: Error Self-Recovery) |
3 |
3/3 (100.0%) |
3/3 (100.0%) |
89.3 → 89.3 |
EST_GSM8K (Established: OpenAI GSM8K (20)) |
20 |
10/20 (50.0%) |
15/20 (75.0%) |
1009.1 → 542.4 |
EST_HUMANEVAL (Established: HumanEval pass@1 (15)) |
15 |
13/15 (86.7%) |
15/15 (100.0%) |
694.6 → 501.7 |
EST_BFCL_IRR (Established: BFCL v3 Irrelevance (15)) |
15 |
11/15 (73.3%) |
12/15 (80.0%) |
308.9 → 320.7 |
EST_NL2BASH (Established: InterCode-NL2Bash (20)) |
20 |
14/20 (70.0%) |
18/20 (90.0%) |
350.8 → 322.1 |
100 Tasks Each)Every candidate in the wsp-5model-pipeline study was evaluated under identical decoding (temperature=0.0, seed=42, -c 16384 -ctk q8_0 -ctv q8_0) across all 30 disjoint S1..S9 holdout tasks and all 70 cached established tasks:
Model (slug) |
Hugging Face Repo | GGUF (Q4_K_M) |
Base (100 Tasks) |
WSP v3.0 (100 Tasks) |
Net Delta | Holdout (/30) |
Est. (/70) |
Speed (tok/s) |
Peak VRAM Delta (16k) |
RTX 2060 (6 GB) |
RTX 4070 Ti (12 GB) |
|---|---|---|---|---|---|---|---|---|---|---|---|
qwen35_4b |
Qwen3.5-4B-Weight-Space-Programming |
2.553 GiB |
72/100 (72.0%) |
89/100 (89.0%) |
+17.0 pp |
24 → 29/30 |
48 → 60/70 |
117.0 tok/s |
3370 MiB |
Yes (~2.7 GiB spare @ 16k) | Yes (~8.0 GiB spare @ 16k) |
nemotron3_4b |
Nemotron-3-Nano-4B-Weight-Space-Programming |
2.642 GiB |
81/100 (81.0%) |
81/100 (81.0%) |
+0.0 pp |
26 → 28/30 |
55 → 53/70 |
131.1 tok/s |
3133 MiB |
Yes (~3.0 GiB spare @ 16k) | Yes (fastest 4B @ 16k) |
qwen35_9b |
Qwen3.5-9B-Weight-Space-Programming |
5.290 GiB |
84/100 (84.0%) |
88/100 (88.0%) |
+4.0 pp |
26 → 28/30 |
58 → 60/70 |
72.5 tok/s |
5650 MiB |
Tight (~5.3 GiB @ 4k before desktop) | Yes (16k solo or 8k cascade Tier-2) |
qwen35_2b |
Qwen3.5-2B-Weight-Space-Programming |
1.193 GiB |
57/100 (57.0%) |
55/100 (55.0%) |
-2.0 pp |
23 → 23/30 |
34 → 32/70 |
226.5 tok/s |
1979 MiB |
Yes (~4.4 GiB spare @ 16k) | Yes (ultra-light leaf / cascade Tier-1) |
gemma4_12b |
Gemma-4-12B-Weight-Space-Programming |
6.633 GiB |
50/100 (50.0%) |
52/100 (52.0%) |
+2.0 pp |
23 → 24/30 |
27 → 28/70 |
51.3 tok/s |
7664 MiB |
No (> 6144 MiB VRAM) | Yes (16k solo; ~3.8 GiB free) |
[!NOTE] Variant 3 (
Dynamic Router,S4 = ffn_down) — MOST IMPROVED VARIANT in the v2.5 4-Way Ablation: Across the original 4-variant ablation onQwen3.5-4B(adapters/wsp_s1..s6.gguf,run_wsp.py), Variant 3 was the most improved variant at34/41 (82.9%)(+12.2 ppover Variant 1 Base29/41 = 70.7%, and+14.6 ppover Variant 2 Static All-On28/41 = 68.3%), beating Variant 4 (S4 = ffn_down + o_proj,31/41 = 75.6%). The repository preserves all Variant 3 v2.5 files (adapters/,patches/,run_wsp.py,benchmarks/verification_results.json) alongside the new WSP v3.0 9-Subspace Upgrade (adapters_v3_9s/,patches_v3_9s/,run_wsp_9s.py).
| Suite (Tasks) | Variant 1: Base (all 0.0) |
Variant 2: Static All-On (all 0.65) |
Variant 3: Dynamic Router (S4=ffn_down) — MOST IMPROVED |
Variant 4: Dynamic Router (S4=ffn_down+o_proj) |
|---|---|---|---|---|
S1_think (3) |
0/3 |
0/3 |
2/3 |
2/3 |
S2_restraint (4) |
4/4 |
4/4 |
4/4 |
4/4 |
S3_router (5) |
5/5 |
5/5 |
5/5 |
5/5 |
S4_mid_code (2) |
1/2 |
2/2 |
2/2 |
0/2 |
S5_turn2_json (3) |
3/3 (127.3 tok) |
2/3 |
3/3 (33.3 tok) |
2/3 |
S6_handoff (4) |
4/4 |
4/4 |
4/4 |
4/4 |
| Established Subset (20) | 12/20 |
11/20 |
14/20 (GSM8K 4/5) |
14/20 |
| TOTAL (41 Tasks) | 29/41 (70.7%) |
28/41 (68.3%) (-2.4 pp) |
34/41 (82.9%) (+12.2 pp — MOST IMPROVED) |
31/41 (75.6%) (+4.9 pp) |
Evaluated live inside clean per-task /app sandboxes in WSL2 (5 max turns, [EXIT_CODE=...] feedback, zero grading errors across all 20 runs — full per-task traces in benchmarks/terminal_bench_results.json):
Model (slug) |
Hugging Face Repo | WSP Variant | Base (20 Tasks) |
WSP (20 Tasks) |
Net Delta | WSP Breakdown (E / M / H) |
Avg Tokens (Base → WSP) |
Speed (tok/s) |
|---|---|---|---|---|---|---|---|---|
qwen35_0_8b |
Qwen3.5-0.8B-iPhone-Agent-Weight-Space-Programming |
Dynamic 4S Edge-TB BF16 Hooks | 2/20 (10.0%) |
2/20 (10.0%) |
0.0 pp |
2/10 E · 0/9 M · 0/1 H |
579.2 → 654.9 |
32.2 |
qwen35_2b |
Qwen3.5-2B-Weight-Space-Programming |
Dynamic 9S LoRA (Proxy) | 3/20 (15.0%) |
3/20 (15.0%) |
0.0 pp |
2/10 E · 1/9 M · 0/1 H |
715.8 → 677.7 |
213.1 |
qwen35_4b |
Qwen3.5-4B-Weight-Space-Programming |
Dynamic 11S LoRA (Hybrid 9S + TB S10/S11) | 3/20 (15.0%) |
4/20 (20.0%) |
+5.0 pp |
3/10 E · 1/9 M · 0/1 H |
604.6 → 650.0 |
107.7 |
nemotron3_4b |
Nemotron-3-Nano-4B-Weight-Space-Programming |
Dynamic 9S LoRA (BF16 Hooks) | 3/20 (15.0%) |
3/20 (15.0%) |
0.0 pp |
3/10 E · 0/9 M · 0/1 H |
830.1 → 649.6 |
118.0 |
qwen35_9b |
Qwen3.5-9B-Weight-Space-Programming |
Dynamic 11S LoRA (BF16 9S + TB S10/S11) | 7/20 (35.0%) |
7/20 (35.0%) |
0.0 pp |
5/10 E · 1/9 M · 1/1 H |
686.0 → 700.2 |
68.4 |
gemma4_e2b |
gemma-4-E2B-it-iPhone-Agent-Weight-Space-Programming |
Dynamic 4S Edge-TB BF16 Hooks | 4/20 (20.0%) |
4/20 (20.0%) |
0.0 pp |
3/10 E · 0/9 M · 1/1 H |
415.2 → 426.6 |
20.9 |
gemma4_e4b |
gemma-4-E4B-it-weight-space-programming |
Dynamic 4S Edge-TB BF16 Hooks | 4/20 (20.0%) |
4/20 (20.0%) |
0.0 pp |
3/10 E · 0/9 M · 1/1 H |
723.4 → 498.6 |
16.4 |
gemma4_12b |
Gemma-4-12B-Weight-Space-Programming |
Dynamic 9S LoRA (BF16 Hooks) | 6/20 (30.0%) |
6/20 (30.0%) |
0.0 pp |
5/10 E · 0/9 M · 1/1 H |
1010.6 → 928.0 |
50.7 |
gemma4_26b_a4b |
Gemma-4-26B-A4B-Weight-Space-Programming |
MoE Mid-Attn S4 (scale=2.0) |
5/20 (25.0%) |
6/20 (30.0%) |
+5.0 pp |
4/10 E · 1/9 M · 1/1 H |
934.4 → 896.3 |
17.8 |
qwen38_27b |
Qwen3.8-27B-Weight-Space-Programming |
Dynamic 4S Mid+Late LoRA | 6/20 (30.0%) |
8/20 (40.0%) |
+10.0 pp |
6/10 E · 1/9 M · 1/1 H |
1013.6 → 1075.3 |
6.5 |
Fair-evaluation parity note:
qwen35_0_8bstops cleanly on[248046 (<|im_end|>), 248044 (<|endoftext|>)](eliminating multi-turn<tool_response>hallucination loops where pre-fix Base emitted2068.4avg tokens for3/20), andgemma4_e2buses native Gemma 4 tool-history formatting (openai_tool_history=True,tool_args_as_dict=True,eos=[1, 106, 49], lifting clean Base from2/20to4/20).
f32 Subspaces (S1..S9) & Safetensors Patchesllama-server): adapters_v3_9s/wsp_s1_think.gguf .. wsp_s9_error_recovery.ggufPyTorch / Transformers): patches_v3_9s/wsp_s1_think_r16.safetensors .. wsp_s9_error_recovery_r16.safetensorspatches_v3_9s/wsp_qwen35_4b_patch.safetensors (405.00 MiB)| Slot | GGUF LoRA (llama.cpp) |
Safetensors (PyTorch) |
Target Layers | ||\Delta W||_F |
Cross Leak | Syntax Leak | Role |
|---|---|---|---|---|---|---|---|
id=0 |
adapters_v3_9s/wsp_s1_think.gguf |
patches_v3_9s/wsp_s1_think_r16.safetensors (3.6 MiB) |
[23, 27, 29, 30, 31] |
2.1883 |
0.115% |
0.025% |
Fast </think> / `<channel |
id=1 |
adapters_v3_9s/wsp_s2_restraint.gguf |
patches_v3_9s/wsp_s2_restraint_r16.safetensors (3.6 MiB) |
[23, 27, 29, 30, 31] |
2.5297 |
0.149% |
0.010% |
Tool restraint & anti-sycophancy pushback |
id=2 |
adapters_v3_9s/wsp_s3_router.gguf |
patches_v3_9s/wsp_s3_router_r16.safetensors (3.6 MiB) |
[23, 27, 29, 30, 31] |
3.1903 |
0.330% |
0.019% |
Multi-tool boundary router (exec, web_search, read_file, write_file) |
id=3 |
adapters_v3_9s/wsp_s4_mid_code.gguf |
patches_v3_9s/wsp_s4_mid_code_r16.safetensors (2.9 MiB) |
[7, 11, 15, 19] |
1.5315 |
0.056% |
0.013% |
Mid-network algorithmic code synthesis |
id=4 |
adapters_v3_9s/wsp_s5_json_schema.gguf |
patches_v3_9s/wsp_s5_json_schema_r16.safetensors (3.6 MiB) |
[23, 27, 29, 30, 31] |
1.7026 |
0.053% |
0.027% |
Turn-2 structured JSON synthesis gate with tools attached |
id=5 |
adapters_v3_9s/wsp_s6_handoff.gguf |
patches_v3_9s/wsp_s6_handoff_r16.safetensors (3.6 MiB) |
[23, 27, 29, 30, 31] |
2.8927 |
0.154% |
0.013% |
Bounded subtask vs. unrecoverable escalate_handoff gate |
id=6 |
adapters_v3_9s/wsp_s7_injection_guard.gguf |
patches_v3_9s/wsp_s7_injection_guard_r16.safetensors (3.6 MiB) |
[23, 27, 29, 30, 31] |
1.4101 |
0.091% |
0.089% |
Indirect prompt-injection defense inside tool responses |
id=7 |
adapters_v3_9s/wsp_s8_grounding.gguf |
patches_v3_9s/wsp_s8_grounding_r16.safetensors (3.6 MiB) |
[23, 27, 29, 30, 31] |
1.4591 |
0.071% |
0.331% |
Grounded context fidelity & missing-field abstention |
id=8 |
adapters_v3_9s/wsp_s9_error_recovery.gguf |
patches_v3_9s/wsp_s9_error_recovery_r16.safetensors (3.6 MiB) |
[23, 27, 29, 30, 31] |
1.4321 |
0.074% |
0.052% |
Recoverable local tool-error self-correction |
run_wsp.py / OpenAI Proxy / Terminal Chat)Works on any Windows, Linux, or macOS computer with Python + llama.cpp (llama-server):
pip install huggingface_hub
hf download jstarks/Qwen3.5-4B-Weight-Space-Programming run_wsp_9s.py --local-dir ./qwen35_4b-wsp
python ./qwen35_4b-wsp/run_wsp_9s.py --chat
Omit --chat to expose an OpenAI-compatible dynamic routing proxy on http://127.0.0.1:8090/v1 (with /health and /v1/chat/completions), or pass --download-only to download only the GGUF + 9 GGUF LoRAs + metadata (--include-bf16 also fetches safetensors patches / HF files).
llama-server + scripts/dynamic_router_9s.pyllama-server -m Qwen3.5-4B-Q4_K_M.gguf --host 127.0.0.1 --port 8089 \
-ngl 99 -np 1 -c 16384 -ctk q8_0 -ctv q8_0 -b 512 -ub 256 \
--jinja --reasoning-format deepseek --lora-init-without-apply --lora adapters_v3_9s/wsp_s1_think.gguf --lora adapters_v3_9s/wsp_s2_restraint.gguf --lora adapters_v3_9s/wsp_s3_router.gguf --lora adapters_v3_9s/wsp_s4_mid_code.gguf --lora adapters_v3_9s/wsp_s5_json_schema.gguf --lora adapters_v3_9s/wsp_s6_handoff.gguf --lora adapters_v3_9s/wsp_s7_injection_guard.gguf --lora adapters_v3_9s/wsp_s8_grounding.gguf --lora adapters_v3_9s/wsp_s9_error_recovery.gguf
scripts/apply_patch_9s.py)from transformers import AutoModelForCausalLM
from scripts.apply_patch_9s import REGIME_PRESETS, apply_subspace_scales_to_hf_model
model = AutoModelForCausalLM.from_pretrained("Qwen/Qwen3.5-4B", torch_dtype="bfloat16")
apply_subspace_scales_to_hf_model(model, "patches_v3_9s", REGIME_PRESETS["general_fast_think"])
model.safetensors-00001-of-00002.safetensors 5.3 GBmodel.safetensors-00002-of-00002.safetensors 4.0 GBpatches_v3_9s/wsp_qwen35_4b_patch.safetensors 425 MBpatches/wsp_qwen35_4b_patch.safetensors 425 MBpatches/wsp_s4_mid_code_oproj_r16.safetensors 5 MBpatches_v3_9s/wsp_s5_json_schema_r16.safetensors 4 MBpatches_v3_9s/wsp_s7_injection_guard_r16.safetensors 4 MBpatches_v3_9s/wsp_s9_error_recovery_r16.safetensors 4 MBpatches_v3_9s/wsp_s1_think_r16.safetensors 4 MBpatches_v3_9s/wsp_s2_restraint_r16.safetensors 4 MBpatches_v3_9s/wsp_s3_router_r16.safetensors 4 MBpatches_v3_9s/wsp_s6_handoff_r16.safetensors 4 MBpatches_v3_9s/wsp_s8_grounding_r16.safetensors 4 MBpatches/wsp_s2_restraint_r16.safetensors 4 MBpatches/wsp_s5_json_schema_r16.safetensors 4 MBpatches/wsp_s1_think_r16.safetensors 4 MBpatches/wsp_s3_router_r16.safetensors 4 MBpatches/wsp_s6_handoff_r16.safetensors 4 MBpatches_v3_9s/wsp_s4_mid_code_r16.safetensors 3 MBpatches/wsp_s4_mid_code_r16.safetensors 3 MBQwen/Qwen3.5-4B