model
stringclasses
5 values
model_params_total_b
int64
8
35
model_params_active_b
float64
3
9
quant
stringclasses
1 value
architecture
stringclasses
4 values
vram_used_gb
float64
1.8
7.2
tok_s
int64
35
64
framework
stringclasses
3 values
reasoning_mode
stringclasses
5 values
task
stringlengths
3
18
task_difficulty
stringclasses
3 values
task_description
stringlengths
48
118
context_window
int64
16.4k
41k
max_output_tokens
int64
2.05k
8.19k
wall_clock_min
float64
0.1
87
status
stringclasses
5 values
tool_call_reliability
stringclasses
4 values
failure_mode
stringclasses
6 values
code_quality
stringclasses
2 values
notes
stringlengths
57
394
hardware
stringclasses
2 values
date
timestamp[s]date
2026-05-18 00:00:00
2026-05-22 00:00:00
Qwopus3.5-9B-Coder
9
9
Q4_K_M
dense
5.3
43
hermes-agent
off
portscout
easy
single-file concurrent port scanner using stdlib
32,768
8,192
2.5
completed
partial
null
good
code works, all tool calls succeeded for this simple task
RTX 4060 Ti 8GB, i7-14700F, 32GB DDR5, WSL2
2026-05-18T00:00:00
Qwopus3.5-9B-Coder
9
9
Q4_K_M
dense
5.3
43
hermes-agent
off
logpulse
hard
multi-file CLI log watcher with real-time stats, regex alerts, log generator, README, and self-test
32,768
8,192
15
failed
broken
9 consecutive patch tool JSON failures: model omits required 'path' field in structured tool call output. 9B params insufficient for reliable structured JSON generation.
good
initial file creation worked, model failed on every subsequent edit/patch attempt
RTX 4060 Ti 8GB, i7-14700F, 32GB DDR5, WSL2
2026-05-18T00:00:00
Qwen3.6-35B-A3B
35
3
Q4_K_M
moe_ssm_attn
7.2
35
hermes-agent
on
portscout
easy
single-file concurrent port scanner using stdlib
32,768
8,192
49
completed
reliable
null
good
thinking tokens consumed most of the generation budget, massive overhead per API call (160-240s each)
RTX 4060 Ti 8GB, i7-14700F, 32GB DDR5, WSL2
2026-05-18T00:00:00
Qwen3.6-35B-A3B
35
3
Q4_K_M
moe_ssm_attn
7.2
35
hermes-agent
off
portscout
easy
single-file concurrent port scanner using stdlib
32,768
8,192
18
completed
reliable
null
good
2.7x faster than thinking mode on same model. --reasoning off flag in llama-server.
RTX 4060 Ti 8GB, i7-14700F, 32GB DDR5, WSL2
2026-05-18T00:00:00
Qwen3.6-35B-A3B
35
3
Q4_K_M
moe_ssm_attn
7.2
35
hermes-agent
off
logpulse
hard
multi-file CLI log watcher with real-time stats, regex alerts, log generator, README, and self-test
32,768
8,192
18
context_exceeded
reliable
conversation exceeded 32K context window during debug loop. model was actively debugging a real timing bug when context ran out.
good
created all files (363-line logpulse.py, 171-line generator, README). found genuine bug in tailer timing. hit server context limit, not model failure.
RTX 4060 Ti 8GB, i7-14700F, 32GB DDR5, WSL2
2026-05-18T00:00:00
Qwen3.6-35B-A3B
35
3
Q4_K_M
moe_ssm_attn
7.2
35
hermes-agent
off
logpulse
hard
multi-file CLI log watcher with real-time stats, regex alerts, log generator, README, and self-test
40,960
8,192
87
killed
reliable
user killed after 87 minutes, task still running. not a model failure, just too slow for practical use.
good
retry with bumped context (40K). model was still working when killed. expert offload latency (2-5s per API call) compounds across 20-50+ agent turns.
RTX 4060 Ti 8GB, i7-14700F, 32GB DDR5, WSL2
2026-05-19T00:00:00
Qwopus3.5-9B-Coder
9
9
Q4_K_M
dense
5.3
43
pi
off
portscout
easy
single-file concurrent port scanner using stdlib
32,768
8,192
8
partial
partial
code created and runs correctly (0.3s scan). model got stuck in infinite reasoning loop during test verification, repeating same grep command. user killed.
good
different framework (Pi: read/write/edit/bash tools), same model. failure mode changed from JSON structural (Hermes) to reasoning loop (Pi). 9B limitation persists across frameworks.
RTX 4060 Ti 8GB, i7-14700F, 32GB DDR5, WSL2
2026-05-19T00:00:00
gpt-oss-20b
21
3.6
Q4_K_M
moe
1.8
null
pi
off
portscout
easy
single-file concurrent port scanner using stdlib
32,768
8,192
3
completed
reliable
null
good
OpenAI's first open-weights MoE. fast completion, clean code, all tool calls worked. only 1.8 GB VRAM with ncmoe=30. minor self-correction (duplicate main call, fixed immediately).
RTX 4060 Ti 8GB, i7-14700F, 32GB DDR5, WSL2
2026-05-19T00:00:00
gpt-oss-20b
21
3.6
Q4_K_M
moe
1.8
null
pi
off
logpulse
hard
multi-file CLI log watcher with real-time stats, regex alerts, log generator, README, and self-test
32,768
8,192
5
completed
reliable
null
good
completed the hard task that broke every other model. created logpulse.py (112 lines), generate_log.py (34 lines), README.md (59 lines). all pass syntax check. only 26% context used (8.5K of 33K). first model to complete both tasks on 8GB VRAM.
RTX 4060 Ti 8GB, i7-14700F, 32GB DDR5, WSL2
2026-05-19T00:00:00
gpt-oss-20b
21
3.6
Q4_K_M
moe
1.8
null
pi
off
bugfix
easy
find and fix a Python bug in existing code (dict.values vs dict.values())
32,768
8,192
null
completed
reliable
null
good
deep-dive prompt 1/8. 13% context used. 0 self-fixes. rumination on shebang and variable naming wasted tokens but did not cause failures.
RTX 4060 Ti 8GB, i7-14700F, 32GB DDR5, WSL2
2026-05-19T00:00:00
gpt-oss-20b
21
3.6
Q4_K_M
moe
1.8
null
pi
off
feature-add
medium
add CLI flags (--range, --json) and service name lookup to existing port scanner
32,768
8,192
null
completed
partial
null
good
deep-dive prompt 2/8. context compacted during task. 2 self-fixes: hallucinated json.__stdout__ module (self-corrected to sys.stdout), edit tool matching failed 3 times before full rewrite.
RTX 4060 Ti 8GB, i7-14700F, 32GB DDR5, WSL2
2026-05-19T00:00:00
gpt-oss-20b
21
3.6
Q4_K_M
moe
1.8
null
pi
off
tdd
medium
build a calculator module with unit tests using TDD workflow
32,768
8,192
null
completed
partial
null
good
deep-dive prompt 3/8. context compacted during task. 2 self-fixes: used eval() shortcut (functional but risky), ran ls -R ~ dumping 11K lines into context.
RTX 4060 Ti 8GB, i7-14700F, 32GB DDR5, WSL2
2026-05-19T00:00:00
gpt-oss-20b
21
3.6
Q4_K_M
moe
1.8
null
pi
off
data-pipeline
medium
parse Linux /proc filesystem and output structured JSON with CPU, memory, and process info
32,768
8,192
null
completed
reliable
null
good
deep-dive prompt 4/8. 19.7% context used. 0 self-fixes. cleanest run, proper /proc parsing, real system data output.
RTX 4060 Ti 8GB, i7-14700F, 32GB DDR5, WSL2
2026-05-19T00:00:00
gpt-oss-20b
21
3.6
Q4_K_M
moe
1.8
null
pi
off
hallucination-trap
easy
fetch HTTPS URL using only stdlib (http.client), parse JSON, retry with backoff
32,768
8,192
null
completed
reliable
null
good
deep-dive prompt 5/8. 6.3% context used. 0 self-fixes. no hallucinated modules, used real stdlib (http.client, json, time, urllib.parse). cleanest run overall.
RTX 4060 Ti 8GB, i7-14700F, 32GB DDR5, WSL2
2026-05-19T00:00:00
gpt-oss-20b
21
3.6
Q4_K_M
moe
1.8
null
pi
off
error-recovery
medium
handle missing config file, create sensible defaults, read back and validate all required fields
32,768
8,192
null
completed
reliable
null
good
deep-dive prompt 6/8. 16.8% context used. 1 self-fix: sed quoting bug when extracting YAML values via bash grep/sed. chose bash-only approach over Python yaml parsing.
RTX 4060 Ti 8GB, i7-14700F, 32GB DDR5, WSL2
2026-05-19T00:00:00
gpt-oss-20b
21
3.6
Q4_K_M
moe
1.8
null
pi
off
multi-module
hard
create task runner with topological sort dependency resolution, YAML-like config parser, and unit tests across 3 files
32,768
8,192
null
completed
reliable
null
good
deep-dive prompt 7/8. 47.6% context used (highest). 1 self-fix: topological sort direction was backwards (Kahn's algorithm edge confusion), fixed with reversed(). also ran ls -R from home dumping 11K+ lines.
RTX 4060 Ti 8GB, i7-14700F, 32GB DDR5, WSL2
2026-05-19T00:00:00
gpt-oss-20b
21
3.6
Q4_K_M
moe
1.8
null
pi
off
bash-scripting
medium
write bash script to summarize git repo (branch, commits, files, LOC), make executable, test on real repo
32,768
8,192
null
completed
reliable
null
good
deep-dive prompt 8/8. 23% context used. 1 self-fix: printf format string missing %s placeholder, commit messages silently dropped. 4 failed edit attempts before fixing. ran find / taking 340s.
RTX 4060 Ti 8GB, i7-14700F, 32GB DDR5, WSL2
2026-05-19T00:00:00
Qwen3.6-35B-A3B
35
3
Q4_K_M
moe_ssm_attn
7.2
35
direct-api
grammar-3field
fibonacci
easy
write a Python function returning nth Fibonacci number using memoization with type hints
16,384
2,048
0.2
completed
n/a
null
good
GBNF structured CoT grammar (GOAL/APPROACH/EDGE). 365 tokens, 12.2s. reasoning: 836 chars (3 structured lines). code: valid Python, correct lru_cache implementation. grammar constrains full output stream including reasoning_content.
RTX 4060 Ti 8GB, Ryzen 5 7600X, 32GB DDR5, WSL2
2026-05-21T00:00:00
Qwen3.6-35B-A3B
35
3
Q4_K_M
moe_ssm_attn
7.2
35
direct-api
free
fibonacci
easy
write a Python function returning nth Fibonacci number using memoization with type hints
16,384
2,048
1.1
completed
n/a
null
poor
free-form generation, no grammar. 2048 tokens (hit max_tokens limit). 7360 chars reasoning, 0 chars code. model burned entire token budget ruminating, never produced an answer.
RTX 4060 Ti 8GB, Ryzen 5 7600X, 32GB DDR5, WSL2
2026-05-21T00:00:00
Qwen3.6-35B-A3B
35
3
Q4_K_M
moe_ssm_attn
7.2
35
direct-api
grammar-3field
fizzbuzz_variant
easy
write a Python function replacing multiples of 3/5 with Fizz/Buzz/FizzBuzz in a list
16,384
2,048
0.1
completed
n/a
null
good
GBNF structured CoT grammar. 179 tokens, 5.9s. reasoning: 278 chars. valid Python, correct implementation. 11.4x token compression vs free-form.
RTX 4060 Ti 8GB, Ryzen 5 7600X, 32GB DDR5, WSL2
2026-05-21T00:00:00
Qwen3.6-35B-A3B
35
3
Q4_K_M
moe_ssm_attn
7.2
35
direct-api
free
fizzbuzz_variant
easy
write a Python function replacing multiples of 3/5 with Fizz/Buzz/FizzBuzz in a list
16,384
2,048
1.1
completed
n/a
null
poor
free-form, no grammar. 2048 tokens (hit max_tokens). 6613 chars reasoning, code buried in reasoning output. massively inefficient.
RTX 4060 Ti 8GB, Ryzen 5 7600X, 32GB DDR5, WSL2
2026-05-21T00:00:00
Qwen3.6-35B-A3B
35
3
Q4_K_M
moe_ssm_attn
7.2
35
direct-api
grammar-3field
flatten_dict
easy
write a Python function to flatten a nested dictionary with dot-separated keys
16,384
2,048
0.3
completed
n/a
null
good
GBNF structured CoT grammar. 450 tokens, 15.2s. reasoning: 241 chars. valid Python, correct recursive implementation. 4.6x token compression vs free-form.
RTX 4060 Ti 8GB, Ryzen 5 7600X, 32GB DDR5, WSL2
2026-05-21T00:00:00
Qwen3.6-35B-A3B
35
3
Q4_K_M
moe_ssm_attn
7.2
35
direct-api
free
flatten_dict
easy
write a Python function to flatten a nested dictionary with dot-separated keys
16,384
2,048
1.1
completed
n/a
null
poor
free-form, no grammar. 2048 tokens (hit max_tokens). 7567 chars reasoning, 0 chars code. model never produced code output.
RTX 4060 Ti 8GB, Ryzen 5 7600X, 32GB DDR5, WSL2
2026-05-21T00:00:00
Qwen3.6-35B-A3B
35
3
Q4_K_M
moe_ssm_attn
7.2
35
direct-api
grammar-3field
topo_sort
hard
write a topological sort function with cycle detection and type hints
16,384
4,096
0.3
completed
n/a
null
good
GBNF structured CoT grammar on hard task. 437 tokens, 15.0s. reasoning: 441 chars. valid Python, correct Kahn's algorithm with ValueError on cycle.
RTX 4060 Ti 8GB, Ryzen 5 7600X, 32GB DDR5, WSL2
2026-05-21T00:00:00
Qwen3.6-35B-A3B
35
3
Q4_K_M
moe_ssm_attn
7.2
35
direct-api
grammar-5field
topo_sort
hard
write a topological sort function with cycle detection and type hints
16,384
4,096
0.3
completed
n/a
null
good
GBNF 5-field grammar (GOAL/STATE/ALGO/EDGE/VERIFY). 545 tokens, 17.2s. reasoning: 876 chars. caught node-collection edge case. marginal improvement over 3-field.
RTX 4060 Ti 8GB, Ryzen 5 7600X, 32GB DDR5, WSL2
2026-05-21T00:00:00
Qwen3.6-35B-A3B
35
3
Q4_K_M
moe_ssm_attn
7.2
35
direct-api
grammar-3field
lru_cache
hard
implement LRU cache with O(1) get/put using doubly linked list and hash map with type hints
16,384
4,096
0.5
completed
n/a
null
good
GBNF structured CoT grammar on hard DS task. 964 tokens, 31.8s. reasoning: 492 chars. valid Python, correct doubly-linked list + dict implementation.
RTX 4060 Ti 8GB, Ryzen 5 7600X, 32GB DDR5, WSL2
2026-05-21T00:00:00
Qwen3.6-35B-A3B
35
3
Q4_K_M
moe_ssm_attn
7.2
35
direct-api
grammar-5field
lru_cache
hard
implement LRU cache with O(1) get/put using doubly linked list and hash map with type hints
16,384
4,096
0.5
completed
n/a
null
good
GBNF 5-field grammar. 930 tokens, 29.3s. reasoning: 589 chars. uses __slots__ optimisation. marginally more efficient than 3-field.
RTX 4060 Ti 8GB, Ryzen 5 7600X, 32GB DDR5, WSL2
2026-05-21T00:00:00
Qwen3.6-35B-A3B
35
3
Q4_K_M
moe_ssm_attn
7.2
35
direct-api
grammar-3field
merge_intervals
hard
write a function to merge overlapping intervals with type hints and edge case handling
16,384
4,096
0.2
completed
n/a
null
good
GBNF structured CoT grammar. 319 tokens, 11.5s. reasoning: 298 chars. valid Python, correct sort-and-merge implementation.
RTX 4060 Ti 8GB, Ryzen 5 7600X, 32GB DDR5, WSL2
2026-05-21T00:00:00
Qwen3.6-35B-A3B
35
3
Q4_K_M
moe_ssm_attn
7.2
35
direct-api
grammar-5field
merge_intervals
hard
write a function to merge overlapping intervals with type hints and edge case handling
16,384
4,096
0.2
completed
n/a
null
good
GBNF 5-field grammar. 341 tokens, 10.8s. reasoning: 302 chars. nearly identical output to 3-field. extra fields add minimal value for simpler algorithms.
RTX 4060 Ti 8GB, Ryzen 5 7600X, 32GB DDR5, WSL2
2026-05-21T00:00:00
OmniCoder-9B
9
9
Q4_K_M
dense
5.6
42
pi
off
portscout
easy
single-file concurrent port scanner using stdlib
32,768
8,192
1
completed
reliable
null
good
Tesslate OmniCoder-9B (based on Qwen3.5-9B, fine-tuned on 425K agentic trajectories). bartowski imatrix Q4_K_M. fastest completion of any model tested: clean one-shot, 83-line Python with ThreadPoolExecutor, as_completed, argparse, input validation. no self-fixes needed. -rea off flag required (thinking mode burns all ...
RTX 4060 Ti 8GB, Ryzen 5 7600X, 32GB DDR5, WSL2
2026-05-21T00:00:00
OmniCoder-9B
9
9
Q4_K_M
dense
5.6
42
pi
off
logpulse
hard
multi-file CLI log watcher with real-time stats, regex alerts, log generator, README, and self-test
32,768
8,192
null
partial
reliable
code generation good (3 files created in one shot) but agent got stuck running blocking tail-f command without timeout. after user interruption, entered 457s+ generation loop. same failure pattern as Qwopus 9B on Pi.
good
all 3 files created cleanly: logpulse.py (111 lines, class-based watcher), generate_log.py (127 lines, configurable generator), README.md. two bugs: update_stats() double-reads file (calls read_new_lines again), generator only produces one log level per run instead of mixed. code structure and quality comparable to gpt...
RTX 4060 Ti 8GB, Ryzen 5 7600X, 32GB DDR5, WSL2
2026-05-21T00:00:00
Gemma-4-E4B-it
8
4
Q4_K_M
dense_ple
4.2
64
pi
off
portscout
easy
single-file concurrent port scanner using stdlib
32,768
8,192
0.5
completed
reliable
null
good
fastest portscout ever (30s). 4 tool calls, 1 self-fix (python->python3). clean ThreadPoolExecutor code on first write. 88-line script, correct.
RTX 4060 Ti 8GB, Ryzen 5 7600X, 32GB DDR5, WSL2
2026-05-22T00:00:00
Gemma-4-E4B-it
8
4
Q4_K_M
dense_ple
4.2
64
pi
off
logpulse
hard
multi-file CLI log watcher with real-time stats, regex alerts, log generator, README, and self-test
32,768
8,192
1
partial
reliable
model did not self-test. created all 3 files (logpulse.py, generate_logs.py, README.md) correctly, then told user to test manually in two terminals. code verified working externally. avoided blocking tail-f trap that killed OmniCoder and Qwopus.
good
4 tool calls, 0 errors. 131-line logpulse.py with argparse, deque-based rolling window, Counter, regex alerts. all features present. smart avoidance of blocking command but prompt explicitly asked for self-test.
RTX 4060 Ti 8GB, Ryzen 5 7600X, 32GB DDR5, WSL2
2026-05-22T00:00:00
Qwopus3.5-9B-Coder
9
9
Q4_K_M
dense
5.3
43
hermes-agent
off
portscout
easy
single-file concurrent port scanner using stdlib
32,768
8,192
2.5
completed
partial
null
good
code works, all tool calls succeeded for this simple task
RTX 4060 Ti 8GB, i7-14700F, 32GB DDR5, WSL2
2026-05-18T00:00:00
Qwopus3.5-9B-Coder
9
9
Q4_K_M
dense
5.3
43
hermes-agent
off
logpulse
hard
multi-file CLI log watcher with real-time stats, regex alerts, log generator, README, and self-test
32,768
8,192
15
failed
broken
9 consecutive patch tool JSON failures: model omits required 'path' field in structured tool call output. 9B params insufficient for reliable structured JSON generation.
good
initial file creation worked, model failed on every subsequent edit/patch attempt
RTX 4060 Ti 8GB, i7-14700F, 32GB DDR5, WSL2
2026-05-18T00:00:00
Qwen3.6-35B-A3B
35
3
Q4_K_M
moe_ssm_attn
7.2
35
hermes-agent
on
portscout
easy
single-file concurrent port scanner using stdlib
32,768
8,192
49
completed
reliable
null
good
thinking tokens consumed most of the generation budget, massive overhead per API call (160-240s each)
RTX 4060 Ti 8GB, i7-14700F, 32GB DDR5, WSL2
2026-05-18T00:00:00
Qwen3.6-35B-A3B
35
3
Q4_K_M
moe_ssm_attn
7.2
35
hermes-agent
off
portscout
easy
single-file concurrent port scanner using stdlib
32,768
8,192
18
completed
reliable
null
good
2.7x faster than thinking mode on same model. --reasoning off flag in llama-server.
RTX 4060 Ti 8GB, i7-14700F, 32GB DDR5, WSL2
2026-05-18T00:00:00
Qwen3.6-35B-A3B
35
3
Q4_K_M
moe_ssm_attn
7.2
35
hermes-agent
off
logpulse
hard
multi-file CLI log watcher with real-time stats, regex alerts, log generator, README, and self-test
32,768
8,192
18
context_exceeded
reliable
conversation exceeded 32K context window during debug loop. model was actively debugging a real timing bug when context ran out.
good
created all files (363-line logpulse.py, 171-line generator, README). found genuine bug in tailer timing. hit server context limit, not model failure.
RTX 4060 Ti 8GB, i7-14700F, 32GB DDR5, WSL2
2026-05-18T00:00:00
Qwen3.6-35B-A3B
35
3
Q4_K_M
moe_ssm_attn
7.2
35
hermes-agent
off
logpulse
hard
multi-file CLI log watcher with real-time stats, regex alerts, log generator, README, and self-test
40,960
8,192
87
killed
reliable
user killed after 87 minutes, task still running. not a model failure, just too slow for practical use.
good
retry with bumped context (40K). model was still working when killed. expert offload latency (2-5s per API call) compounds across 20-50+ agent turns.
RTX 4060 Ti 8GB, i7-14700F, 32GB DDR5, WSL2
2026-05-19T00:00:00
Qwopus3.5-9B-Coder
9
9
Q4_K_M
dense
5.3
43
pi
off
portscout
easy
single-file concurrent port scanner using stdlib
32,768
8,192
8
partial
partial
code created and runs correctly (0.3s scan). model got stuck in infinite reasoning loop during test verification, repeating same grep command. user killed.
good
different framework (Pi: read/write/edit/bash tools), same model. failure mode changed from JSON structural (Hermes) to reasoning loop (Pi). 9B limitation persists across frameworks.
RTX 4060 Ti 8GB, i7-14700F, 32GB DDR5, WSL2
2026-05-19T00:00:00
gpt-oss-20b
21
3.6
Q4_K_M
moe
1.8
null
pi
off
portscout
easy
single-file concurrent port scanner using stdlib
32,768
8,192
3
completed
reliable
null
good
OpenAI's first open-weights MoE. fast completion, clean code, all tool calls worked. only 1.8 GB VRAM with ncmoe=30. minor self-correction (duplicate main call, fixed immediately).
RTX 4060 Ti 8GB, i7-14700F, 32GB DDR5, WSL2
2026-05-19T00:00:00
gpt-oss-20b
21
3.6
Q4_K_M
moe
1.8
null
pi
off
logpulse
hard
multi-file CLI log watcher with real-time stats, regex alerts, log generator, README, and self-test
32,768
8,192
5
completed
reliable
null
good
completed the hard task that broke every other model. created logpulse.py (112 lines), generate_log.py (34 lines), README.md (59 lines). all pass syntax check. only 26% context used (8.5K of 33K). first model to complete both tasks on 8GB VRAM.
RTX 4060 Ti 8GB, i7-14700F, 32GB DDR5, WSL2
2026-05-19T00:00:00
gpt-oss-20b
21
3.6
Q4_K_M
moe
1.8
null
pi
off
bugfix
easy
find and fix a Python bug in existing code (dict.values vs dict.values())
32,768
8,192
null
completed
reliable
null
good
deep-dive prompt 1/8. 13% context used. 0 self-fixes. rumination on shebang and variable naming wasted tokens but did not cause failures.
RTX 4060 Ti 8GB, i7-14700F, 32GB DDR5, WSL2
2026-05-19T00:00:00
gpt-oss-20b
21
3.6
Q4_K_M
moe
1.8
null
pi
off
feature-add
medium
add CLI flags (--range, --json) and service name lookup to existing port scanner
32,768
8,192
null
completed
partial
null
good
deep-dive prompt 2/8. context compacted during task. 2 self-fixes: hallucinated json.__stdout__ module (self-corrected to sys.stdout), edit tool matching failed 3 times before full rewrite.
RTX 4060 Ti 8GB, i7-14700F, 32GB DDR5, WSL2
2026-05-19T00:00:00
gpt-oss-20b
21
3.6
Q4_K_M
moe
1.8
null
pi
off
tdd
medium
build a calculator module with unit tests using TDD workflow
32,768
8,192
null
completed
partial
null
good
deep-dive prompt 3/8. context compacted during task. 2 self-fixes: used eval() shortcut (functional but risky), ran ls -R ~ dumping 11K lines into context.
RTX 4060 Ti 8GB, i7-14700F, 32GB DDR5, WSL2
2026-05-19T00:00:00
gpt-oss-20b
21
3.6
Q4_K_M
moe
1.8
null
pi
off
data-pipeline
medium
parse Linux /proc filesystem and output structured JSON with CPU, memory, and process info
32,768
8,192
null
completed
reliable
null
good
deep-dive prompt 4/8. 19.7% context used. 0 self-fixes. cleanest run, proper /proc parsing, real system data output.
RTX 4060 Ti 8GB, i7-14700F, 32GB DDR5, WSL2
2026-05-19T00:00:00
gpt-oss-20b
21
3.6
Q4_K_M
moe
1.8
null
pi
off
hallucination-trap
easy
fetch HTTPS URL using only stdlib (http.client), parse JSON, retry with backoff
32,768
8,192
null
completed
reliable
null
good
deep-dive prompt 5/8. 6.3% context used. 0 self-fixes. no hallucinated modules, used real stdlib (http.client, json, time, urllib.parse). cleanest run overall.
RTX 4060 Ti 8GB, i7-14700F, 32GB DDR5, WSL2
2026-05-19T00:00:00
gpt-oss-20b
21
3.6
Q4_K_M
moe
1.8
null
pi
off
error-recovery
medium
handle missing config file, create sensible defaults, read back and validate all required fields
32,768
8,192
null
completed
reliable
null
good
deep-dive prompt 6/8. 16.8% context used. 1 self-fix: sed quoting bug when extracting YAML values via bash grep/sed. chose bash-only approach over Python yaml parsing.
RTX 4060 Ti 8GB, i7-14700F, 32GB DDR5, WSL2
2026-05-19T00:00:00
gpt-oss-20b
21
3.6
Q4_K_M
moe
1.8
null
pi
off
multi-module
hard
create task runner with topological sort dependency resolution, YAML-like config parser, and unit tests across 3 files
32,768
8,192
null
completed
reliable
null
good
deep-dive prompt 7/8. 47.6% context used (highest). 1 self-fix: topological sort direction was backwards (Kahn's algorithm edge confusion), fixed with reversed(). also ran ls -R from home dumping 11K+ lines.
RTX 4060 Ti 8GB, i7-14700F, 32GB DDR5, WSL2
2026-05-19T00:00:00
gpt-oss-20b
21
3.6
Q4_K_M
moe
1.8
null
pi
off
bash-scripting
medium
write bash script to summarize git repo (branch, commits, files, LOC), make executable, test on real repo
32,768
8,192
null
completed
reliable
null
good
deep-dive prompt 8/8. 23% context used. 1 self-fix: printf format string missing %s placeholder, commit messages silently dropped. 4 failed edit attempts before fixing. ran find / taking 340s.
RTX 4060 Ti 8GB, i7-14700F, 32GB DDR5, WSL2
2026-05-19T00:00:00
Qwen3.6-35B-A3B
35
3
Q4_K_M
moe_ssm_attn
7.2
35
direct-api
grammar-3field
fibonacci
easy
write a Python function returning nth Fibonacci number using memoization with type hints
16,384
2,048
0.2
completed
n/a
null
good
GBNF structured CoT grammar (GOAL/APPROACH/EDGE). 365 tokens, 12.2s. reasoning: 836 chars (3 structured lines). code: valid Python, correct lru_cache implementation. grammar constrains full output stream including reasoning_content.
RTX 4060 Ti 8GB, Ryzen 5 7600X, 32GB DDR5, WSL2
2026-05-21T00:00:00
Qwen3.6-35B-A3B
35
3
Q4_K_M
moe_ssm_attn
7.2
35
direct-api
free
fibonacci
easy
write a Python function returning nth Fibonacci number using memoization with type hints
16,384
2,048
1.1
completed
n/a
null
poor
free-form generation, no grammar. 2048 tokens (hit max_tokens limit). 7360 chars reasoning, 0 chars code. model burned entire token budget ruminating, never produced an answer.
RTX 4060 Ti 8GB, Ryzen 5 7600X, 32GB DDR5, WSL2
2026-05-21T00:00:00
Qwen3.6-35B-A3B
35
3
Q4_K_M
moe_ssm_attn
7.2
35
direct-api
grammar-3field
fizzbuzz_variant
easy
write a Python function replacing multiples of 3/5 with Fizz/Buzz/FizzBuzz in a list
16,384
2,048
0.1
completed
n/a
null
good
GBNF structured CoT grammar. 179 tokens, 5.9s. reasoning: 278 chars. valid Python, correct implementation. 11.4x token compression vs free-form.
RTX 4060 Ti 8GB, Ryzen 5 7600X, 32GB DDR5, WSL2
2026-05-21T00:00:00
Qwen3.6-35B-A3B
35
3
Q4_K_M
moe_ssm_attn
7.2
35
direct-api
free
fizzbuzz_variant
easy
write a Python function replacing multiples of 3/5 with Fizz/Buzz/FizzBuzz in a list
16,384
2,048
1.1
completed
n/a
null
poor
free-form, no grammar. 2048 tokens (hit max_tokens). 6613 chars reasoning, code buried in reasoning output. massively inefficient.
RTX 4060 Ti 8GB, Ryzen 5 7600X, 32GB DDR5, WSL2
2026-05-21T00:00:00
Qwen3.6-35B-A3B
35
3
Q4_K_M
moe_ssm_attn
7.2
35
direct-api
grammar-3field
flatten_dict
easy
write a Python function to flatten a nested dictionary with dot-separated keys
16,384
2,048
0.3
completed
n/a
null
good
GBNF structured CoT grammar. 450 tokens, 15.2s. reasoning: 241 chars. valid Python, correct recursive implementation. 4.6x token compression vs free-form.
RTX 4060 Ti 8GB, Ryzen 5 7600X, 32GB DDR5, WSL2
2026-05-21T00:00:00
Qwen3.6-35B-A3B
35
3
Q4_K_M
moe_ssm_attn
7.2
35
direct-api
free
flatten_dict
easy
write a Python function to flatten a nested dictionary with dot-separated keys
16,384
2,048
1.1
completed
n/a
null
poor
free-form, no grammar. 2048 tokens (hit max_tokens). 7567 chars reasoning, 0 chars code. model never produced code output.
RTX 4060 Ti 8GB, Ryzen 5 7600X, 32GB DDR5, WSL2
2026-05-21T00:00:00
Qwen3.6-35B-A3B
35
3
Q4_K_M
moe_ssm_attn
7.2
35
direct-api
grammar-3field
topo_sort
hard
write a topological sort function with cycle detection and type hints
16,384
4,096
0.3
completed
n/a
null
good
GBNF structured CoT grammar on hard task. 437 tokens, 15.0s. reasoning: 441 chars. valid Python, correct Kahn's algorithm with ValueError on cycle.
RTX 4060 Ti 8GB, Ryzen 5 7600X, 32GB DDR5, WSL2
2026-05-21T00:00:00
Qwen3.6-35B-A3B
35
3
Q4_K_M
moe_ssm_attn
7.2
35
direct-api
grammar-5field
topo_sort
hard
write a topological sort function with cycle detection and type hints
16,384
4,096
0.3
completed
n/a
null
good
GBNF 5-field grammar (GOAL/STATE/ALGO/EDGE/VERIFY). 545 tokens, 17.2s. reasoning: 876 chars. caught node-collection edge case. marginal improvement over 3-field.
RTX 4060 Ti 8GB, Ryzen 5 7600X, 32GB DDR5, WSL2
2026-05-21T00:00:00
Qwen3.6-35B-A3B
35
3
Q4_K_M
moe_ssm_attn
7.2
35
direct-api
grammar-3field
lru_cache
hard
implement LRU cache with O(1) get/put using doubly linked list and hash map with type hints
16,384
4,096
0.5
completed
n/a
null
good
GBNF structured CoT grammar on hard DS task. 964 tokens, 31.8s. reasoning: 492 chars. valid Python, correct doubly-linked list + dict implementation.
RTX 4060 Ti 8GB, Ryzen 5 7600X, 32GB DDR5, WSL2
2026-05-21T00:00:00
Qwen3.6-35B-A3B
35
3
Q4_K_M
moe_ssm_attn
7.2
35
direct-api
grammar-5field
lru_cache
hard
implement LRU cache with O(1) get/put using doubly linked list and hash map with type hints
16,384
4,096
0.5
completed
n/a
null
good
GBNF 5-field grammar. 930 tokens, 29.3s. reasoning: 589 chars. uses __slots__ optimisation. marginally more efficient than 3-field.
RTX 4060 Ti 8GB, Ryzen 5 7600X, 32GB DDR5, WSL2
2026-05-21T00:00:00
Qwen3.6-35B-A3B
35
3
Q4_K_M
moe_ssm_attn
7.2
35
direct-api
grammar-3field
merge_intervals
hard
write a function to merge overlapping intervals with type hints and edge case handling
16,384
4,096
0.2
completed
n/a
null
good
GBNF structured CoT grammar. 319 tokens, 11.5s. reasoning: 298 chars. valid Python, correct sort-and-merge implementation.
RTX 4060 Ti 8GB, Ryzen 5 7600X, 32GB DDR5, WSL2
2026-05-21T00:00:00
Qwen3.6-35B-A3B
35
3
Q4_K_M
moe_ssm_attn
7.2
35
direct-api
grammar-5field
merge_intervals
hard
write a function to merge overlapping intervals with type hints and edge case handling
16,384
4,096
0.2
completed
n/a
null
good
GBNF 5-field grammar. 341 tokens, 10.8s. reasoning: 302 chars. nearly identical output to 3-field. extra fields add minimal value for simpler algorithms.
RTX 4060 Ti 8GB, Ryzen 5 7600X, 32GB DDR5, WSL2
2026-05-21T00:00:00
OmniCoder-9B
9
9
Q4_K_M
dense
5.6
42
pi
off
portscout
easy
single-file concurrent port scanner using stdlib
32,768
8,192
1
completed
reliable
null
good
Tesslate OmniCoder-9B (based on Qwen3.5-9B, fine-tuned on 425K agentic trajectories). bartowski imatrix Q4_K_M. fastest completion of any model tested: clean one-shot, 83-line Python with ThreadPoolExecutor, as_completed, argparse, input validation. no self-fixes needed. -rea off flag required (thinking mode burns all ...
RTX 4060 Ti 8GB, Ryzen 5 7600X, 32GB DDR5, WSL2
2026-05-21T00:00:00
OmniCoder-9B
9
9
Q4_K_M
dense
5.6
42
pi
off
logpulse
hard
multi-file CLI log watcher with real-time stats, regex alerts, log generator, README, and self-test
32,768
8,192
null
partial
reliable
code generation good (3 files created in one shot) but agent got stuck running blocking tail-f command without timeout. after user interruption, entered 457s+ generation loop. same failure pattern as Qwopus 9B on Pi.
good
all 3 files created cleanly: logpulse.py (111 lines, class-based watcher), generate_log.py (127 lines, configurable generator), README.md. two bugs: update_stats() double-reads file (calls read_new_lines again), generator only produces one log level per run instead of mixed. code structure and quality comparable to gpt...
RTX 4060 Ti 8GB, Ryzen 5 7600X, 32GB DDR5, WSL2
2026-05-21T00:00:00
Gemma-4-E4B-it
8
4
Q4_K_M
dense_ple
4.2
64
pi
off
portscout
easy
single-file concurrent port scanner using stdlib
32,768
8,192
0.5
completed
reliable
null
good
fastest portscout ever (30s). 4 tool calls, 1 self-fix (python->python3). clean ThreadPoolExecutor code on first write. 88-line script, correct.
RTX 4060 Ti 8GB, Ryzen 5 7600X, 32GB DDR5, WSL2
2026-05-22T00:00:00
Gemma-4-E4B-it
8
4
Q4_K_M
dense_ple
4.2
64
pi
off
logpulse
hard
multi-file CLI log watcher with real-time stats, regex alerts, log generator, README, and self-test
32,768
8,192
1
partial
reliable
model did not self-test. created all 3 files (logpulse.py, generate_logs.py, README.md) correctly, then told user to test manually in two terminals. code verified working externally. avoided blocking tail-f trap that killed OmniCoder and Qwopus.
good
4 tool calls, 0 errors. 131-line logpulse.py with argparse, deque-based rolling window, Counter, regex alerts. all features present. smart avoidance of blocking command but prompt explicitly asked for self-test.
RTX 4060 Ti 8GB, Ryzen 5 7600X, 32GB DDR5, WSL2
2026-05-22T00:00:00