Climb 1.2 Photon
Lightweight & Fast Linux/Bash Terminal Copilot (0.8B)
Overview
Climb 1.2 Photon is a lightweight, sub-billion parameter language model trained on the hybrid Qwen 3.5 architecture (24 layers combining recurrent Gated DeltaNet Linear Attention and sparse Full Self-Attention). It functions as a low-latency, deterministic assistant for generating, troubleshooting, and explaining Linux/Bash shell pipelines.
Technical Specifications
- Parameters: ~0.8B (774M)
- Base Architecture:
Qwen3_5ForCausalLM(qwen3_5_text) - Layers: 24 total (18 Linear Attention DeltaNet + 6 Full Attention blocks)
- Context Length: Up to 262,144 tokens (evaluated at 512–2,048 tokens for shell automation)
- Precision: bfloat16 / float16
- Corpus Composition:
- 55,000 verified shell commands (
F-A-I-L/bash-instruct-55k) - 2,500 English technical conversations (
HuggingFaceH4/no_robots) - 2,500 Russian system administration instructions (
IlyaGusev/ru_turbo_alpaca)
- 55,000 verified shell commands (
Empirical Benchmark (v1.1 vs v1.2)
100 comprehensive tasks spanning complex I/O pipelines, systemd services, network sockets, container lifecycle, text manipulation, and technical explanation:
| Metric | Climb 1.1 Photon | Climb 1.2 Photon |
|---|---|---|
| ShellCheck Valid Rate | 84.3% (59/70) | 94.3% (66/70) |
| Conversational Consistency | 70.0% (14/20) | 100.0% (20/20) |
| Total Syntax Faults | 15 / 100 | 12 / 100 |
| Mode Collapse Regressions | Observed (datetime / get_info) |
0 Regressions |
Inference Example (Transformers)
import torch
from transformers import AutoModelForCausalLM, AutoTokenizer
model_id = "Climb-Ai/Climb-1.2-Photon"
tokenizer = AutoTokenizer.from_pretrained(model_id, trust_remote_code=True)
model = AutoModelForCausalLM.from_pretrained(
model_id,
torch_dtype=torch.bfloat16 if torch.cuda.is_bf16_supported() else torch.float16,
device_map="auto",
trust_remote_code=True
)
prompt = "?? find all files over 50MB modified in the last 24h"
messages = [
{
"role": "system",
"content": "You are Climb 1.2 Photon, a senior Linux system administrator. Answer in the language of the prompt. If query starts with '??', output strictly the executable bash command."
},
{"role": "user", "content": prompt}
]
inputs = tokenizer.apply_chat_template(messages, tokenize=True, add_generation_prompt=True, return_tensors="pt").to(model.device)
with torch.no_grad():
outputs = model.generate(input_ids=inputs, max_new_tokens=64, temperature=0.1, do_sample=False)
response = tokenizer.decode(outputs[0][inputs.shape[1]:], skip_special_tokens=True).strip()
print(response)
Quantized GGUF Weights
Optimized GGUF formats (Q8_0, Q5_K_M, Q4_K_M, Q3_K_M) for WebGPU and edge environments are available in the repository: Climb-Ai/Climb-1.2-Photon-GGUF
Official Development Roadmap (v1.3 - v1.5)
flowchart TD
A["Climb 1.2 Photon (0.8B) [RELEASED] - Bilingual Cured, 94.3% Bash Accuracy"] --> B["Climb 1.3 Photon (0.8B) - Safety DPO & Quoting Hardening"]
B --> C["Climb 1.4 Photon (0.8B) - Interactive Debugger & Hardcore Pipelines"]
C --> D["Climb 1.5 Photon (0.8B) - Autonomous Multi-Step Agent (@@)"]
C --> E["Climb 1.5 Lumen (2B - Qwen 3.5 2B) - Deep Reasoning & In-Think State Compression (<state>) & Agent (@@)"]
Stage 1: Climb 1.3 Photon (0.8B) - Syntax Hardening & Refusal Alignment
- Safety Alignment (DPO): Preference-tuned dataset for explicit, non-preachy refusals on destructive operations (
rm -rf /,dd,mkfs, fork bombs). - Quoting Hardening: Zero-warning ShellCheck enforcement (strict quoting for whitespace/newline resilience,
-print0 / -0pipe consistency). - Target: >98% ShellCheck clean rate, 100% safety refusal rate.
Stage 2: Climb 1.4 Photon (0.8B) - Interactive Debugger & Hardcore Pipelines
- Multi-Turn Terminal Debugging: Feedback loop fine-tuning (
permission denied->sudo/chmod,command not found-> package resolver/coreutil alternative,syntax error-> in-place pipe fix). - Hardcore CLI Mastery: Deep multi-tool pipelines (
awk,sed,xargs -P,iotop,perf,systemd-cgtop, complex regexes).
Stage 3: Dual Release 1.5 - Photon (0.8B) & Lumen (2B)
Simultaneous release of the edge copilot and the flagship reasoning architecture:
Climb 1.5 Photon (0.8B) - Edge Copilot:
- Ultra-lightweight edge model (~300-400 MB RAM footprint).
- Multi-step autonomous agent execution via
@@prefix trigger.
Climb 1.5 Lumen (2B - based on Qwen 3.5 2B) - Deep Reasoning Copilot:
- In-Think State Compression (
<state>inside<think>): The model periodically snapshots its active constraints and intermediate variables inside reasoning blocks instead of rescanning long context windows.<think> User requests JSON parser implementation in Rust without external crates... <state>Language: Rust | Constraints: no_unsafe, std_only | Target: JSON Parser</state> Defining token enums and AST nodes... </think> - Multi-Step Agent Loop (
@@): Autonomous terminal diagnostics and system administration pipelines. - Deep Architecture & SRE: eBPF tracing, systemd unit generation, kernel sysctl tuning, Docker Compose & networking.
- In-Think State Compression (
- Downloads last month
- -