felkf/SIQ-1-35B-oQ6-fp16
<div align="center"> <img src="assets/siq_logo.png" width="78%"/> </div>
<div align="center"> 🤗 <a href="https://huggingface.co/AlexWortega/SIQ-1-35B"><b>Model</b></a>   |    📦 <a href="https://huggingface.co/AlexWortega/SIQ-1-35B/tree/main/gguf"><b>GGUF</b></a>   |    📚 <a href="https://huggingface.co/datasets/AlexWortega/karp-autoresearch-distill"><b>Dataset</b></a>   |    🪽 <a href="https://huggingface.co/spaces/AlexWortega/hermes-agent-zerogpu"><b>Demo</b></a> </div>
SIQ-1-tiny-35b 🪽
A tiny universal agent — autoresearch, coding, reasoning.
SIQ-1-tiny-35b is a tiny MoE — 35B total but only ~3B active per token — distilled to be a strong universal agent: equally at home running autonomous ML research (autoresearch), writing and debugging code, tool-use / agentic workflows, and hard reasoning. Despite its 3B active footprint it matches or beats much larger peers on core reasoning, sycophancy-resistance, and agentic coding — at a lower token cost.
Autoresearch duel (head-to-head)
<div align="center"> <img src="assets/autoresearch_duel.png" width="92%"/> </div>
In a controlled three-way autoresearch test on `openai/parameter-golf` — each model driving the same Pi-Agent edit train_gpt.py -> train (300s) -> eval val_bpb -> keep/revert loop on its own 1xA6000 for 2h — SIQ-1-tiny-35b reached val_bpb 1.767 (12 experiments, full 2h), neck-and-neck with Claude Opus 4.8 (~1.76) and far ahead of GLM-5.2 (2.078). GLM stagnated on the baseline — its only hypothesis was "add depth" (which hurt the metric) and it stopped emitting actions after ~65 min; SIQ instead climbed via LR-schedule and capacity edits (warmdown 1200->800, matrixlr 0.04->0.05, ...). *(valbpb on a single A6000 is not comparable to the official 8xH100 leaderboard; this is the relative head-to-head under identical conditions.)*
It is the winning arm of a controlled SFT / RFT / DPO / offline-GRPO post-training study on Qwen3.6-35B-A3B: ppo on the judge-top-half wins both ideation quality and agentic ability.
Performance
On the full 198-question GPQA-Diamond — all models served as Q4_K_M GGUF, greedy (temp 0), identical harness — SIQ-1-tiny-35b is *Pareto-best: the highest accuracy and the fewest tokens* (figure below). A 3B-active model edging out a full 35B base and Nex-N2-mini, while spending fewer tokens per question.
<div align="center"> <img src="assets/gpqa_pareto.png" width="78%"/> </div>
bf16 + tuned harness scores higher (90.2 GPQA); the Q4 row is the apples-to-apples co-measured comparison shown in the figure. Terminal-Bench 2.1 (Harbor, terminus-2, k=5) is in progress.
BullshitBench v2 — pushback vs. sycophancy
<div align="center"> <img src="assets/bullshit_bar.png" width="68%"/> </div>
Score 0–2 (Clear Pushback = 2 / Partial = 1 / Accepted = 0). Panel: claude-sonnet-4.6 + gpt-5.2 + gemini-3.1-pro (mean), judge sees the final answer only (CoT stripped); no system prompt, temp 0.7. | model | avg /2 | Clear Pushback | Partial | Accepted | |---|---|---|---|---| | SIQ-1-tiny-35b (high/think) | 1.047 | 45 | 17 | 38 | | Nex-N2-Pro (free) | 1.040 | 33 | 43 | 24 |
A tie on the mean, but different profiles: SIQ is polarized (cleanly exposes the BS 45× or fully buys it 38×); Nex hedges (rarely fully accepts, but rarely pushes back hard either — mostly Partial). Reference (official bullshit-benchmark, different panel, n=55, not co-measured): Opus 4.8 ≈ 1.96, GPT-5.5 ≈ 0.92.
Reasoning modes & system prompts
Qwen3-format hybrid reasoning, toggled per request via chat_template_kwargs.enable_thinking:
Reasoning effort is a trained control — Reasoning effort: low | medium | high in the system prompt scales the chain length (high for hard reasoning). For objective reasoning use greedy (temp 0) — it beats temp 0.7 by ~8 pts.
Copy-paste system prompts:
1 · Hard reasoning — greedy + high effort
Reasoning effort: high. Think step by step inside <think>...</think>, then give the final answer.2 · Autoresearch ideator — propose a train.py edit to cut val_bpb
Reasoning effort: high. You are an autoresearch ideator.
Given the current train.py and its measured val_bpb under a fixed compute budget, propose ONE concrete,
high-impact edit that should reduce val_bpb. Reason inside <think>...</think>, then output:
- a one-line hypothesis,
- the edit as a minimal unified diff,
- the expected effect and how to verify it.3 · Fast / instruction-following — no-think
(no system prompt; set enable_thinking=false — the model answers directly, no <think> block)Usage
🪽 Try it now (no install): hosted ZeroGPU demo → **AlexWortega/hermes-agent-zerogpu**
llama.cpp (GGUF — single 48 GB GPU; Q4KM ≈ 21 GB)
These are the exact flags we serve with:
docker run -d --gpus all --network host -v /models:/m ghcr.io/ggml-org/llama.cpp:server-cuda \
-m /m/SIQ-1-35B.Q4_K_M.gguf --alias SIQ-1-tiny-35b \
-ngl 99 -c 131072 -np 4 --jinja --host 0.0.0.0 --port 8080--jinjarequired (Qwen3 chat template →<think>+ tool tags; enablesenable_thinking).-ngl 99all layers on GPU;-c 131072total context split across-np 4slots (≈32k/slot — agentic loops need the headroom). Drop to-c 65536if you only do short reasoning. OpenAI-compatible on:8080.
sglang (bf16 safetensors — e.g. 2× 48 GB)
python -m sglang.launch_server \
--model-path AlexWortega/SIQ-1-35B \
--tp 2 --context-length 131072 \
--reasoning-parser qwen3 --tool-call-parser qwen3 \
--host 0.0.0.0 --port 8080Call it (OpenAI-compatible) — these are the params we run
from openai import OpenAI
client = OpenAI(base_url="http://localhost:8080/v1", api_key="x")
r = client.chat.completions.create(
model="SIQ-1-tiny-35b",
messages=[{"role": "system", "content": "Reasoning effort: high"},
{"role": "user", "content": "..."}],
temperature=0.0, top_p=0.95, top_k=40, # greedy (temp 0) for reasoning
extra_body={"chat_template_kwargs": {"enable_thinking": True}}) # False → no-thinkSampling: reasoning → temperature 0 (greedy); general/creative → temp 0.7, topp 0.95, topk 40. Files: merged bf16 *.safetensors + GGUF Q4_K_M / Q5_K_M / Q8_0 (+ MTP f16).