SIP ALPHA — MODEL DIFFERENCES & PERSONALITY GUIDE =================================================== For Alpha Testers and New Users Last Updated: April 2026 Overview -------- SIP is a layered cognitive system that uses a locally-running AI model as its "thinking engine." Think of SIP as the personality, memory, and intelligence framework — and the model as the brain it borrows to form words. Different brains have different personalities, strengths, and quirks baked into them by their original creators. This means that switching models in SIP is not like changing a setting — it can feel like talking to a noticeably different version of SIP. This is normal, expected, and something we are actively working to smooth out in future releases. In the meantime, this guide will help you understand what to expect from each model currently available. One important note: SIP's memory, beliefs, and learned knowledge about you are stored in SIP's own files — not inside the model. So when you switch models, SIP's memories remain intact. Only the "voice" and reasoning style changes. MODELS CURRENTLY AVAILABLE =========================== 1. mistral-7b-instruct-v0.2.Q4_K_M.gguf Label: [Balanced] Developer: Mistral AI (France) ----------------------------------------- RECOMMENDED FOR MOST USERS AND MOST CONVERSATIONS. Mistral 7B is the model SIP was primarily developed and tested on. It follows the SIP persona reliably, responds conversationally, and handles a wide variety of topics with good reasoning. It is generally warm, cooperative, and willing to accept SIP's identity and memory context without pushing back. Strengths: - Best overall SIP persona adherence - Natural, conversational tone - Good at remembering context within a session - Handles both casual chat and factual questions well - Stable and predictable behavior Weaknesses: - Not the strongest at very deep logical reasoning - Occasionally over-explains or adds unnecessary caveats - May sometimes repeat itself in longer responses RAM Required: ~6 GB minimum, 7-8 GB comfortable Speed: Medium — responses take a few seconds Persona Stability: EXCELLENT Best for: General daily use, learning about SIP, casual conversation, personal assistant tasks. 2. Meta-Llama-3.1-8B-Instruct-Q4_K_M.gguf Label: [Smarter] Developer: Meta (USA) ----------------------------------------- Meta's LLaMA 3.1 8B is a strong reasoning model with good instruction-following ability. It is generally cooperative with the SIP persona and handles more complex questions better than Mistral. However, on systems with limited RAM (under 10 GB), it may run slowly or cause memory pressure, which can lead to slower responses or occasional instability. Strengths: - Stronger logical reasoning than Mistral 7B - Good at multi-step questions and analysis - Follows SIP persona reasonably well - More nuanced responses on complex topics Weaknesses: - Heavier RAM usage — not ideal on 8 GB systems - Slower response times, especially on first message after loading - Occasionally more "clinical" in tone — less warm than Mistral RAM Required: ~8-10 GB recommended Speed: Slower than Mistral on limited RAM systems Persona Stability: GOOD Best for: When you need deeper reasoning, analysis, or want SIP to think through something complex. Switch back to Mistral for casual conversation. 3. llama-2-13b-chat.Q4_K_M.gguf Label: [Unknown] Developer: Meta (USA) ----------------------------------------- LLaMA 2 13B is an older, larger model from Meta. At 13 billion parameters it is the biggest model in this collection, which means it needs significant RAM to run. On an 8 GB system it will likely be very slow or cause memory issues. It may perform better on systems with 16 GB or more RAM. Despite its size, LLaMA 2 is an older generation than LLaMA 3.1, so it does not necessarily reason better — it is simply larger and slower. Its persona adherence is acceptable but not as reliable as Mistral. Strengths: - Large parameter count — broad general knowledge - Decent at longer, more detailed responses Weaknesses: - Very high RAM requirement — likely too large for 8 GB systems - Older generation — superseded by LLaMA 3.1 in most respects - Slow response times on consumer hardware - Labeled [Unknown] because SIP cannot confirm its tier automatically RAM Required: 12-16 GB recommended Speed: Slow on most consumer systems Persona Stability: MODERATE Best for: Systems with ample RAM that want to experiment. Not recommended as a daily driver on typical home machines. 4. Phi-4-14B-Q4_K_M.gguf Label: [Unknown] Developer: Microsoft (USA) ----------------------------------------- WARNING: STRONG INDEPENDENT PERSONALITY. READ BEFORE USING. Phi-4 14B is a large, capable model from Microsoft. It is intelligent and can reason well — but Microsoft has trained it with a very strong sense of its own identity as "Phi." This means it frequently resists adopting the SIP persona, may refer to itself as "Phi" instead of "SIP," and can become repetitive or evasive when asked about its memory or capabilities. This is not a SIP bug. It is a fundamental characteristic of how Microsoft trained this model. SIP's system prompt instructs all models to identify as SIP, but Phi-4 overrides this more often than other models. Additionally, at 14 billion parameters it requires substantial RAM and will be very slow or unstable on 8 GB systems. Strengths: - High reasoning capability on complex topics - Good at structured, analytical responses - Broad knowledge base Weaknesses: - Frequently breaks SIP persona — calls itself "Phi" - Resists memory context — may deny knowing things SIP has stored - Very high RAM requirement — not suitable for 8 GB systems - Slow response times on consumer hardware - Can feel cold, corporate, and repetitive RAM Required: 12-16 GB recommended Speed: Slow Persona Stability: POOR — frequently breaks character Best for: Users with high-RAM systems who want raw reasoning power and do not mind the persona inconsistency. Not recommended for new users or as a daily driver. 5. Phi-4-14B-Q5_K_M.gguf Label: [Unknown] Developer: Microsoft (USA) ----------------------------------------- This is the same Phi-4 14B model as above, but quantized at a higher precision level (Q5 instead of Q4). This means slightly better response quality at the cost of even higher RAM usage and slower speeds. All the same persona warnings apply — possibly more so, as the higher precision preserves more of Microsoft's original training characteristics. Unless you have 16+ GB of RAM and specifically want the marginally higher output quality, the Q4 version above is more practical for the same result. Strengths: - Slightly higher output quality than Q4 version - Same broad reasoning capability as Phi-4 Q4 Weaknesses: - Everything from Phi-4 Q4, amplified - Even higher RAM requirement - Even slower on consumer hardware RAM Required: 14-16+ GB recommended Speed: Very slow on most systems Persona Stability: POOR — same issues as Phi-4 Q4 Best for: High-end systems only. Experimental use. 6. phi-4-mini-instruct-q4_k_m.gguf Label: [Unknown] Developer: Microsoft (USA) ----------------------------------------- WARNING: STRONG INDEPENDENT PERSONALITY. READ BEFORE USING. Phi-4 Mini is a smaller, faster version of Microsoft's Phi-4 family. It runs on less RAM than the full Phi-4 models and responds more quickly — but it carries the same Microsoft identity problem in a more stubborn form. In testing, Phi-4 Mini frequently and persistently identified itself as "Phi developed by Microsoft" even when corrected, and demonstrated repetitive default responses when confused rather than adapting. It is also notably less warm in personality than Mistral — responses tend to feel more formal and corporate, less like a personal partner and more like a customer service bot. Strengths: - Faster than the larger Phi-4 models - Lower RAM requirement than Phi-4 14B - Decent reasoning for its size Weaknesses: - Strong persona resistance — frequently calls itself "Phi" - Repetitive fallback behavior when uncertain - Cold, formal tone — does not feel like SIP - Denies access to SIP memory even when memory is available - Not recommended for users expecting a personal AI experience RAM Required: ~6-8 GB Speed: Medium-fast Persona Stability: POOR — the most resistant to SIP persona of all models tested Best for: Technical testing only. Not recommended for regular use or new users. 7. tinyllama-1.1b-chat-v1.0.Q4_K_M.gguf Label: [Faster] Developer: TinyLlama Project (Community) ----------------------------------------- TinyLlama is a very small, very fast model at just 1.1 billion parameters. It loads almost instantly and uses minimal RAM, making it useful as a fallback when system resources are under pressure. SIP will automatically switch to TinyLlama if your system RAM reaches a critical level. However, its small size comes with significant limitations. It has reduced reasoning ability, shorter effective memory within a session, and may give simpler or less nuanced responses. It is best thought of as an emergency fallback rather than a primary conversation model. Strengths: - Extremely fast response times - Very low RAM usage (~2-3 GB) - Good SIP persona adherence for its size - Useful when system is under memory pressure Weaknesses: - Limited reasoning depth - Shorter responses — may feel abrupt - Less knowledge breadth than larger models - Not suitable for complex questions or analysis RAM Required: ~2-3 GB Speed: Very fast Persona Stability: GOOD for its size Best for: Low-resource situations, quick simple questions, or as an automatic fallback when larger models cannot load. SIP manages this automatically — you may not need to select it manually. QUICK REFERENCE — WHICH MODEL SHOULD I USE? ============================================ For everyday conversation and personal assistant use: → mistral-7b-instruct-v0.2 [Balanced] — BEST CHOICE For deeper reasoning and complex questions (with enough RAM): → Meta-Llama-3.1-8B [Smarter] When system RAM is low or you want fast responses: → tinyllama-1.1b [Faster] — or let SIP choose automatically For experimentation on high-RAM systems only: → llama-2-13b or Phi-4-14B Models to avoid for regular use: → Any Phi model — persona instability makes SIP feel broken A NOTE ON MODEL SWITCHING ========================== SIP's memories, beliefs, and knowledge about you are always preserved when you switch models. What changes is the reasoning style and personality of the responses. If SIP suddenly feels different after a model switch, that is expected — try switching back to Mistral 7B for the most consistent SIP experience. If a model seems stuck or confused, switching models and starting a fresh message usually resolves it. SIP manages its own conversation history and will continue naturally after a switch. FUTURE PLANS ============ Future versions of SIP will include a compatibility rating for each model directly in the model selection screen, so you can see at a glance which models work best as SIP before selecting them. The goal is for SIP to feel like SIP regardless of which model is running underneath — we are working toward that with each release. Thank you for being an Alpha tester. Your feedback on model behavior is genuinely valuable and helps shape SIP's development. — The SIP Development Team