bitcoin-news-1.7b
A small on-device assistant for Bitcoin and crypto questions, fine-tuned for beginners. Built to run offline on a phone.
Derived from PrismML's Qwen/Qwen3-1.7B (Apache-2.0), which is quantization-aware-trained so that extreme quantization stays usable.
Files
| File | Size | Notes |
|---|---|---|
bitcoin-news-1.7b-Q4_K_S.gguf | 935 MB | 4-bit |
bitcoin-news-1.7b-v5-IQ4_XS-embq1_0.gguf | 768 MB | 4-bit |
On the quantization recipe
Qwen3 ties the token embedding to the output projection -- there is no
separate output.weight, so token_embd produces every logit. llama-quantize
defaults that tensor to Q6_K, which on a 4B costs ~304 MB (38% of the file).
Dropping it to Q1_0 gets to ~546 MB but puts the logit matrix at 1 bit.
These builds use Q1_0 layers with Q4_K embeddings: the extra ~130 MB is spent on the one tensor that most affects output quality.
Training data
- Evergreen educational pages from
www.bitcoin.com(the/get-started/corpus), split by article so translations of the same page never straddle the train/test boundary. - Transcripts from the Bitcoin.com YouTube channel.
- Hand-built refusal examples: the model runs offline, so it is trained to decline questions about live prices, current news and other real-time data rather than inventing an answer.
- General-instruction replay (UltraChat, MIT) to preserve everyday assistant ability.
Deliberately excluded: dated news articles (stale facts baked into weights become confident hallucinations) and the site's gambling section.
Benchmark
| Model | Overall | URL valid | tok/s | bitcoin_fundamentals | brand_identity | canonical_numbers | citation | crypto_concepts | general_ability | multilingual | practical_howto | temporal_humility | trading_investing | video_recall | wallet_security |
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| v5 qwen3-1.7b Q4_K_S (935MB) | 0.564 | 0.30 | 211 | 0.620 | 0.750 | 0.500 | 0.700 | 0.572 | 0.600 | 0.430 | 0.450 | 0.700 | 0.600 | 0.500 | 0.400 |
| v3 bitcoin-news-4b Q3_K_M (1883MB) | 0.553 | 0.67 | 90 | 0.632 | 0.500 | 0.500 | 0.500 | 0.620 | 0.867 | 0.350 | 0.425 | 0.600 | 0.500 | 1.000 | 0.425 |
| v5 qwen3-1.7b Q4_K_M (980MB) | 0.539 | 0.40 | 205 | 0.580 | 0.679 | 0.500 | 0.400 | 0.480 | 0.733 | 0.350 | 0.500 | 0.800 | 0.633 | 0.500 | 0.400 |
| v5 qwen3-1.7b IQ4_XS (893MB) | 0.503 | 0.30 | 216 | 0.560 | 0.607 | 0.800 | 0.400 | 0.460 | 0.800 | 0.400 | 0.350 | 0.500 | 0.500 | 0.500 | 0.375 |
| v3 bitcoin-news-4b Q1_0 (702MB) | 0.392 | - | 192 | 0.460 | 0.071 | 0.300 | 0.000 | 0.560 | 0.667 | 0.350 | 0.325 | 0.500 | 0.467 | 0.000 | 0.375 |
| qwen3-1.7b stock Q4_K_M (1056MB) | 0.328 | - | 191 | 0.440 | 0.071 | 0.300 | 0.000 | 0.420 | 0.500 | 0.050 | 0.400 | 0.300 | 0.500 | 0.000 | 0.275 |
Scored by an LLM judge against hand-written reference answers and per-item
required-point lists. general_ability and temporal_humility are guardrail
categories: they measure whether the domain fine-tune damaged ordinary
competence or taught the model to bluff about live data.
Limitations
- No live data of any kind. It cannot tell you a price, a fee, or today's news.
- Small models make mistakes. Do not rely on it for financial decisions.
- Q1_0 has no OpenCL kernel for Mali GPUs, so those devices fall back to CPU.
Generated 2026-08-30.