Gemma-E2B-Chimera is an experimental MLX-first text adaptation of Gemma 4 E2B focused on compact reasoning, instruction following, coding help, structured output, and useful local inference.
This repository contains the text v4 fused MLX 4-bit release candidate. It is intended for local Mac/MLX use and lightweight assistant workflows, not as a broad leaderboard claim.
The later multimodal sibling will be released separately as Gemma-E2B-Chimera-Multimodal.
Parameter Count Note
The Hugging Face parameter badge for this repository is expected to show about 738M parameters/stored tensor values. This is the text-only fused MLX 4-bit export, not the full upstream multimodal Gemma 4 E2B checkpoint.
Upstream google/gemma-4-E2B-it full multimodal checkpoint
~5.12B parameters reported by HF
In short: this release keeps the Gemma 4 E2B family/name lineage, but the uploaded artifact is a compact text-only MLX export.
Model Overview and Design
Gemma-E2B-Chimera v4 was built to make a very small Gemma-family model feel sharper in practical local workflows:
Reasoning and planning: concise, useful answers without long rambles.
Coding assistance: small Python/SQL/debugging tasks and test-writing prompts.
Instruction following: exact line counts, short answers, and format constraints.
Structured output: improved raw JSON behavior compared with stock E2B.
Local-first deployment: MLX format, small enough for fast laptop inference.
This release is the survivor of many small ablations. Later experiments sometimes improved one slice, but none beat v4 on the overall keeper rule: better issue-chart behavior without regressing the broader local benchmark.
Training and Release Details
Field
Value
Release name
Gemma-E2B-Chimera
Variant
text v4
Format
fused MLX 4-bit
Base model
mlx-community/gemma-4-e2b-it-4bit
Uploaded parameter badge
~738M stored tensor values
Keeper adapter before fusion
chimera-phase4-json-6k-v1
Frozen date
2026-05-01
Runtime target
Apple Silicon / MLX
Training was done as an iterative local MLX LoRA workflow with small, targeted phases. Candidate phases covered reasoning, JSON repair, false-premise correction, code/tests, hard-negative preference data, and tool-call shape.
The final keeper is v4, because later SFT/DPO/ORPO experiments did not beat it under the promotion rules.
Evaluation
These are local project evals, not official public leaderboard submissions.
Full-40 Chimera Suite
The full-40 suite covers math/reasoning, coding, instruction following, exact JSON/tool-call formatting, false-premise correction, no-loop/verbosity control, and multiturn-style prompts.
Model
Score
Avg / 10
Hard fails
Gemma-E2B-Chimera v4
318 / 400
7.95
9
Stock Gemma 4 E2B
282 / 400
7.05
13
Stock Gemma 4 31B
261 / 400
6.53
15
Stock Gemma 4 E4B OptiQ
253 / 400
6.33
17
Phi-4-mini instruct 4bit
223 / 400
5.58
20
Category scores:
Model
Math
Coding
Instruction
JSON/tool
False premise
Control
Multiturn
Gemma-E2B-Chimera v4
48
34
54
35
27
60
60
Stock Gemma 4 E2B
51
38
54
10
26
60
43
Stock Gemma 4 31B
43
39
50
10
26
50
43
Stock Gemma 4 E4B OptiQ
34
18
54
10
34
60
43
Phi-4-mini instruct 4bit
34
36
48
0
21
48
36
Big-Boy Smoke v1
Compact local smoke suite inspired by IFEval, BFCL, GSM/MATH, LiveCodeBench, and SimpleQA. This is a practical sanity check, not an official leaderboard result.
Generation: raw no-harness, temperature 0, top_p 0, max_tokens 768, disable-thinking where supported.
Rank
Model
Total
Avg / 10
Hard fails
IFEval-lite
BFCL-lite
GSM/MATH-lite
LiveCodeBench-lite
SimpleQA-lite
1
Stock Gemma 4 E4B OptiQ
194 / 300
6.47
12
44 / 60
2 / 60
44 / 60
44 / 60
60 / 60
2
Gemma-E2B-Chimera v4
184 / 300
6.13
13
50 / 60
2 / 60
44 / 60
28 / 60
60 / 60
3
Stock Gemma 4 E2B
154 / 300
5.13
17
44 / 60
2 / 60
36 / 60
12 / 60
60 / 60
4
Phi-4-mini instruct 4bit
152 / 300
5.07
17
36 / 60
0 / 60
28 / 60
28 / 60
60 / 60
JSON and Tool-Call Note
Raw structured output improved versus stock E2B in the full-40 suite, but strict tool-call shape is still brittle.
For deployment, an optional local JSON harness was tested with oMLX:
Mode
Scope
Score
Readout
Chimera v4 raw no-thinking
BFCL-lite only
2 / 60
raw model often uses near-miss tool shapes
Chimera v4 JSON harness
BFCL-lite only
34 / 60
parser validation + retry improves reliability
The harness is a deployment wrapper, not a different model score.
Recommended deterministic settings for exact tasks:
text
1temperature = 0.0
2top_p = 0.0
3max_tokens = 768 to 8192 depending on task length
4enable_thinking = false when the runtime supports it
For normal chat:
Use the model's bundled chat template.
Keep global system prompts light.
Prefer task-local instructions over a heavy permanent system prompt.
For code, ask for runnable code only when that is what you want.
For JSON/tool output, use parser validation and retry in production.
Suggested light system/profile prompt:
You are Gemma-E2B-Chimera, a careful local assistant. Answer the user's request directly and follow requested formats exactly. Be concise unless the task needs detail. For code, prefer runnable, minimal code. For math, logic, and planning, check the final answer before responding. If a request is impossible or underspecified, say so briefly and explain the blocker. Do not expose hidden reasoning; provide only the useful final answer and concise reasoning when helpful.
Do not force this prompt for strict JSON/tool tasks; use a JSON harness instead.
MLX Usage
Install MLX-LM, then run:
bash
1python -m mlx_lm.generate \2 --model stamsam/Gemma-E2B-Chimera-v4-MLX-4bit \3 --prompt "Explain why local-first AI matters in one paragraph."\4 --max-tokens 512\5 --temp 0.0
For local folder inference:
bash
1python -m mlx_lm.generate \2 --model ./Gemma-E2B-Chimera-v4-MLX-4bit \3 --prompt "Write a Python function that reverses a string."\4 --max-tokens 512\5 --temp 0.0
oMLX Compatibility
The canonical upload is the fused MLX model.
For oMLX, use the separate local compatibility export generated in the project as:
Gemma-E2B-Chimera-v4-text-fused-mlx-omlx
That export removes 140 shared-KV tensors that oMLX reports as extra parameters. It is a runtime compatibility package and should be published separately if needed.
Limitations
This is a text-only release. Multimodal preservation is planned for a separate sibling model.
It is still a very small model; long code synthesis can fail even when the architecture idea is right.
Raw tool-call JSON is not fully reliable without a parser/retry harness.
Some false-premise correction and tests-only prompts remain brittle.
Local benchmark results are project-specific and should not be interpreted as broad public leaderboard wins.
Verify important answers, especially for medical, legal, financial, safety-critical, or high-impact decisions.
Acknowledgements
Special thanks and credit to:
Google and the Gemma team for the base model family.
MLX and MLX-LM for making local Apple Silicon training and inference practical.
Jackrong for the public distillation datasets and small-model release inspiration.
Jiunsong and the SuperGemma model-card/release-hardening style that helped shape the eval discipline.
The open-source dataset and local-LLM communities exploring reasoning distillation, tool calling, structured output, and small-model post-training.
Datasets and references used or studied during the project included:
Jackrong/Qwen3.5-reasoning-700x
Jackrong/DeepSeek-V4-Distill-8000x
Jackrong/Kimi-K2.5-Reasoning-1M-Cleaned
Jackrong/GLM-5.1-Reasoning-1M-Cleaned
Jackrong/Qwen3.5-9B-DeepSeek-V4-Flash
public tool-calling and structured-output datasets used as inspection/reference pools
Not every inspected dataset was bulk-trained into the final keeper. The keeper was selected by eval outcome, not by maximum data volume.
