SIQ-1-35B — Public model and demo¶

Development line: project:siq-1-35b · thread public-model-and-demo
Last event: 2026-06-17 · 1 dated since 2026-06-17 · Researched: 2026-09-05 · confidence: medium

What it is¶

SIQ-1-35B — an Apache-2.0 finetune of Qwen/Qwen3.6-35B-A3B for local coding, reasoning, and tool use.

• Thinking and no-think chat modes
• BF16 and GGUF distributions
• Local OpenAI-compatible serving

Development line¶

• 2026-06-17 — SIQ-1-35B public model and agent demo were linked. On 2026-06-17, SIQ-1-35B was linked to a public Hugging Face model page and a Hermes-agent ZeroGPU Space. Autoresearch, parameter-golf, and benchmark repositories were linked alongside it. The links confirm the public files, but not the author's benchmark claims.

What changed¶

2026-06-17 — The model card renamed SIQ-1-35B to “SIQ-1-tiny-35b”. It presents the model as a universal agent for autoresearch, coding, and reasoning. The update added benchmark charts, serving profiles, prompts, and a demo link, without a weight upload recorded that day.

How to use this¶

From 2026-06-17, treat SIQ-1-35B as a public model with an agent demo. Verify capabilities and benchmark claims independently before relying on them.

1. Start with the Q4_K_M GGUF locally: llama serve -hf AlexWortega/SIQ-1-35B:Q4_K_M, then use its localhost OpenAI-compatible endpoint. — https://huggingface.co/AlexWortega/SIQ-1-35B
2. For tool calls and thinking, use the documented llama.cpp profile with --jinja, GPU offload, and a 131072-token context split across four slots. — https://huggingface.co/AlexWortega/SIQ-1-35B
3. For BF16 serving, use the documented SGLang configuration with tensor parallelism 2 plus the Qwen3 reasoning and tool-call parsers. — https://huggingface.co/AlexWortega/SIQ-1-35B
4. Set enable_thinking=true, Reasoning effort: high, and temperature 0 for hard reasoning. Turn thinking off for direct replies. — https://huggingface.co/AlexWortega/SIQ-1-35B
5. Before enabling MTP speculative decoding, fetch the current companion GGUF and confirm it loads with the selected trunk quantization. — https://huggingface.co/AlexWortega/SIQ-1-35B/commit/cb6f95ec76411517babbfa6a854f7f0425313ff4

Best practices¶

• Keep --jinja enabled for llama.cpp: the documented Qwen chat template carries the thinking and tool tags. — https://huggingface.co/AlexWortega/SIQ-1-35B
• Report benchmark results with precision, prompt mode, and sampling settings; the card distinguishes Q4 co-measurement from its BF16 result. — https://huggingface.co/AlexWortega/SIQ-1-35B
• Treat MTP as version-sensitive: the companion was corrected on 2026-06-23 after an incorrect draft layer, and users reported earlier load failures. — https://huggingface.co/AlexWortega/SIQ-1-35B/discussions/2
• Use a local runtime rather than the hosted ZeroGPU demo until it recovers; it returned a missing libcuda.so.1 runtime error when read. — https://huggingface.co/spaces/AlexWortega/hermes-agent-zerogpu
• Do not present the card’s one-A6000, two-hour test as an official Parameter Golf leaderboard result; that project scopes leaderboard submissions to ten minutes on 8×H100. — https://github.com/openai/parameter-golf

Superseded by this¶

• 2026-06-17 — The model is a finetune, not an adapter. The release history explicitly corrects this relationship.
• 2026-06-23 — Earlier vocab-only DFlash and wrong-layer MTP companions are obsolete; the repository replaced them with the Qwen3.6-35B-A3B MTP layer and updated tensor names.

Still unknown¶

• Independent evaluations for SIQ claims are missing. Treat GPQA, autoresearch, coding, and BullshitBench figures as author-reported.
• SIQ-1-35B is the repository name while its 2026-06-17 card calls the same model SIQ-1-tiny-35b. Commit history shows no separate model identities.
• Karpathy autoresearch and OpenAI Parameter Golf use different training and evaluation setups. Sources show no direct training-data lineage to SIQ-1-35B.
• SIQ is tagged as text generation although its Qwen3.6 base is multimodal. The model card does not confirm vision input support.
• The hosted ZeroGPU failure is a hosted-runtime issue, not evidence that local GGUF or BF16 serving fails.

Sources¶

Agent brief¶

• Subject: project:siq-1-35b, thread public-model-and-demo, 1 dated events 2026-06-17 → 2026-06-17.
• Practical note: From 2026-06-17, treat SIQ-1-35B as having a linked public model artifact and agent-demo entry point, while independently verifying capabilities and benchmark claims before relying on them.
• Confidence: medium. Dated supersedes above are the authority for what is obsolete.