Datasets:

---

witcheer

/

local-agentic-coding-bench-8gb-vram-2026-05

like 4

OmniCoder-9B agentic benchmark: fastest code generation, same agent loop problems (2 new rows)

tested Tesslate/OmniCoder-9B (dense 9B, based on Qwen3.5-9B, fine-tuned on 425K+ agentic coding trajectories) on our standard portscout + logpulse tasks via Pi agent framework.

setup

• model: OmniCoder-9B Q4_K_M (bartowski imatrix quant, 5.6 GB)
• server: llama-server (turboquant), -rea off flag mandatory
• framework: Pi coding agent (read/write/edit/bash tools)
• hardware: RTX 4060 Ti 8GB, Ryzen 5 7600X, 32 GB DDR5, WSL2
• decode speed: 42 tok/s (dense, fully in VRAM)

critical note: thinking mode must be disabled

with default settings, OmniCoder-9B burns 100% of tokens on reasoning_content and produces 0 code output. this is inherited from the Qwen3.5-9B base. the -rea off server flag disables thinking mode and routes output to content. without this flag, the model is completely unusable.

portscout (easy): PASS, fastest model tested

OmniCoder produced the cleanest first-shot output of any model: 83-line Python with ThreadPoolExecutor, as_completed, argparse with proper metavar/type hints, input validation for port ranges, and configurable thread count. no self-fixes needed, no rumination, no wasted tokens.

logpulse (hard): PARTIAL, agent stuck

all 3 files were created cleanly in one shot:

• logpulse.py (111 lines): class-based watcher with file position tracking, regex pattern matching, stats (total lines, lines/sec, error count)
• generate_log.py (127 lines): configurable fake log generator with seed, append mode, level selection
• README.md: usage docs with examples

two code bugs:

1. update_stats() calls read_new_lines() again (double-reads, double-counts)
2. generator only produces one log level per run instead of mixing INFO/WARN/ERROR randomly

the real failure was agent loop management:

1. model ran logpulse.py test.log (a blocking tail -f watcher) without timeout or background flag
2. command hung indefinitely waiting for new log lines
3. after user interrupted, model tried to debug but entered a 457s+ generation loop
4. user killed the session

this is the exact same failure pattern as Qwopus 9B on Pi. the model doesn't understand that tail -f style commands block forever and need to be run in background or with a timeout.

key finding

agentic fine-tuning improves code generation quality but does NOT fix agent loop management. OmniCoder's 425K agentic trajectories made it the fastest and cleanest code generator we've tested, but the model still can't plan multi-step agent workflows involving blocking commands. code generation and agent orchestration are separate skills.

gpt-oss-20b remains the winner for end-to-end agentic work. OmniCoder-9B could be a strong choice for single-shot code generation tasks where agent loop complexity is minimal.

data

rows 30-31 in agentic-coding-bench.jsonl. filter by model == "OmniCoder-9B".