Measured effect
8 held-out eval-v3 canonical fail questions, first web_search forced to fail, greedy next tool:
Table with columns: model, next web_search, next open_page| model | next web_search | next open_page |
|---|
| Javi0.3-2B | 0/8 | 8/8 |
| Javi0.4-2B | 7/8 | 1/8 |
This probe isolates the decision the edit targets. In a two-seed harness canary with real tools and sampling, 0.84 chose open_page after a permanently blocked search and completed, same as the base; it used one extra search on a normal lookup. A 0.75 scale flipped 4/4 but looped on 14 searches under a permanent block, so 0.84 is the conservative setting. The edit reaches the canonical open_page tool; in opencode/grok, webfetch/web_fetch after a 403 is unchanged.
Full eval v3
175 held-out seeds, 3 runs, date pin 2026-10-07, cached search results. Javi0.3-2B answers on the same seeds are the baseline. Judges: DeepSeek-V4.1-Flash-1017 and mimo-v26-dual.
Table with columns: metric, Javi0.3-2B, Javi0.4-2B| metric | Javi0.3-2B | Javi0.4-2B |
|---|
| core pass (judge average) | 0.35 [0.30, 0.40] | 0.41 [0.36, 0.45] |
| core pass (deepseek) | 0.48 | 0.58 |
| core pass (mimo) | 0.23 | 0.24 |
| recovery after a tool failure | 0.55 (n=11) | 0.45 (n=10) |
| made-up links | 0.01 | 0.00 |
| loops | 0.04 | 0.03 |
Paired core pass difference: +0.06 [0.02, 0.10] (deepseek +0.10 [0.05, 0.16], mimo +0.02 [-0.02, 0.06]), 95% bootstrap over 175 seeds. The deepseek interval clears zero; the mimo interval crosses zero, so the size of the gain is uncertain. Conflict questions improved (0.33 vs 0.07, n=7) and nodate (0.42 vs 0.17, n=6); premise (0.23 vs 0.29, n=13) and recovery after real failures did not.
Five answer episodes (4 seeds) carry no deepseek verdict because the judge omits the tools or style key; one seed never finished an episode in any of the three runs and the judge skips it, so judge-only rows cover 174 seeds. Core pass stays at 175 seeds with incomplete episodes scored 0. Full scoreboard: data/v2/eval/javi04-vs-03-2b/score.md in the source tree.
Files
Merged bf16 safetensors (one shard, 4.2 GB), tokenizer, processor files and chat template. The vision tower is Qwen3.5-2B's, unchanged; the edit is in the language model only.
Run it
Javi was trained and evaluated with thinking off. Send "chat_template_kwargs": {"enable_thinking": false} and use temperature 0.2, top-p 0.8, top-k 20 so tool calls stay well formed.
GGUF for llama.cpp or lmdeck: Javi0.4-2B-GGUF.
lmdeck
Transformers
from transformers import AutoModelForImageTextToText, AutoProcessor model_id = "javierbits/Javi0.4-2B"processor = AutoProcessor.from_pretrained(model_id)model = AutoModelForImageTextToText.from_pretrained(model_id, dtype="auto", device_map="auto") tools = [{"type": "function", "function": { "name": "web_search", "description": "Search the web.", "parameters": {"type": "object", "properties": {"query": {"type": "string"}}, "required": ["query"]}}}]messages = [{"role": "user", "content": [{"type": "text", "text": "What's the latest stable Rust release?"}]}] inputs = processor.apply_chat_template(messages, tools=tools, add_generation_prompt=True, enable_thinking=False, tokenize=True, return_dict=True, return_tensors="pt").to(model.device)out = model.generate(**inputs, max_new_tokens=512, do_sample=True, temperature=0.2, top_p=0.8, top_k=20)print(processor.decode(out[0][inputs["input_ids"].shape[1]:], skip_special_tokens=True))
The edit
embed_tokens[51187] *= 0.84 # =open_page
Token 51187 is a single token in the Qwen3.5 tokenizer. The row is tied between input and output embeddings, so the output logit for =open_page is reduced.
Links