Model family← Models

Qwen3-Coder-Next

The coding version of Qwen3-Next 80B-A3B. Weaker than its sibling at extracting facts and saying "I don't know", but the data it built in a real task was sound.

Files tested: 1
Scored runs: 3

OverviewWhy we tested it · what happened · should you use it

How this page was made. The tests run automatically on our Mac Studio, and every score here comes straight from those runs. This overview was written by Claude, an AI model, from the lab's own records: the run log, notes kept at the time, and the working sessions where each decision was made. Quotes are Alan's words, with spelling fixed. How these pages are made.

Why we tested it

It is the coding version of Qwen3-Next 80B-A3B, which we already ran every day: the same design, trained for coding, tool use and recovering when a command fails. It uses about 3 billion of its 80 billion parameters for each token and doesn't think before answering.

On 27 August Alan asked: "Qwen coder next is specialised at coding no?" He made it an experiment: "3 models, same family, same quant, same context, let’s learn why they exist and what they’re best at". So it ran beside the Instruct and Thinking versions of Qwen3-Next.

What it took

Unsloth's IQ4_XS on stock llama.cpp, matching the Instruct model's file so every score compares. The file alone is 42.7 GB.

What happened

The coding training cost it some care. Answering from a document with the right source, it scored 76.8% against the Instruct version's 88.8%. When the answer wasn't there, it said so 56.9% of the time against 97.2%. At telling good extractions from bad they were level: 90.0% against 91.7%.
On the document test it never made an answer up. Most of its misses there were incomplete or badly formed.
In real use its work held up. In one of our agent loops it wrote the sample data for a stock-control app: 28,094 records, every link between them valid. We had predicted, from its weak "not there" score, that the data would be broken. It wasn't.
It held a long conversation. The last turn ran with 193,443 tokens in play, at 14.7 tokens a second.
Deleted on a wrong assumption. On 8 September it was deleted to free disk. The deletion list said it had never been tested; its results were filed under another name, and nothing was lost. Alan: "we can revisit coder models later".

What it taught us

Same design, different training, different strengths. The Instruct training bought saying "I don't know"; the coding training traded it for getting things built.
A weak "not there" score didn't predict weak work. That test measures declining under pressure, not building.
Spotting errors didn't change with training. Four Qwen models we compared all scored 86.7% to 91.7% at it, failing the same four cases.
Name results the way you name files. A check by folder name found none of its results.

Should you use it?

If your Mac has room (check the smallest Mac for the file below) and you want a model that builds without stopping to think, its work held up in a real task here. For extracting facts, or anywhere "I don't know" matters, Qwen3-Next-80B was much better. We never ran it on our tool or coding tests, so its strengths there are unmeasured.

The files1 tested · each has its own page

Qwen3-Coder-Next IQ4_XS
Unsloth's IQ4_XS, the same format as its Instruct sibling so the scores compare. Deleted on 8 September.
Unsloth GGUF · Deleted 8 Sep 2026 · 42.7 GB · 3 scored runs · smallest Mac 64 GB

The lab's recordEvery decision, with its date · the model's specification

Decisions, in order

2026-08-27 07:09Z, fetch Coder-Next. Alan "Pull it also"; the assistant's IQ4_XS choice. Alternatives rejected: Q3 quants for the full window, Q4_0/Q4_K_S which exceed the ceiling at 262K.
2026-08-27 07:15Z, pre-registered predictions (assistant, before any Coder data): "Worth stating predictions now, before the data, so they can be wrong." Coder wins coinpusher decisively; Coder loses nullschema (30B coder: 37.5); Coder ties or slightly loses grounding. Scored 08:29 to 08:32Z: nullschema loss confirmed (56.9); grounding "slightly loses" too soft (12 points); coinpusher never run.
2026-08-27 08:25Z, router swap for the suites (evicts Instruct for ~15 minutes) after the gate passed.
2026-08-27 15:07Z, Alan runs Coder in DSH. "running same with qwen coder 80b now". Pre-registered by the assistant: fewer rounds, better build loop, and "worse stub data" (the sharp one). Scored 2026-08-28 15:21Z: the stub-data prediction was wrong.
2026-08-31, left resident and idle. The model sat loaded at 41.55 GB from the DSH session's last turn (Thursday 20:07) for 3 days 6 hours; Alan: "No idea what coder doing, I have been out all weekend". The assistant did not unload it.
2026-09-08 07:47Z and 14:03Z, the deletion list said "zero results". The assistant first noted Coder-Next had zero results, then listed it as "zero results, never once benchmarked". That was wrong: the results exist under qwen3-next-80b-coder. The pre-delete check searched for the directory name and printed qwen3-coder-next 0 result files and 0 runlog entries; the deletion still ran at 14:05:35Z, and the assistant told Alan afterwards that both deleted presets had banked results. Alan, queued at 14:05:11Z: "we can revisit coder models later". Later: "we can regrab a model whenever".
2026-09-08 14:07Z, restore script. RESTORE-deleted-20260908.sh function coder re-downloads unsloth/Qwen3-Coder-Next-GGUF --include "*IQ4_XS*".

What we learned, in the lab's words

From the controlled trio:

Instruct buys abstention; Coder trades it for execution. Run-log scores for this label (variant doc): grounding 76.8, screening 90.0, nullschema 56.9, against Instruct 88.8 / 91.7 / 97.2. Failure signature differs by regime: Coder's grounding failures are mostly partial (11) and malformed (8) with few misattributions (10) and no fabrication; Thinking's are misattribution (16) and budget-malformed (14).
Detection is insensitive to post-training. Screening 86.7 to 91.7 across four checkpoints; every error in every model is a false reject on scr/correct/*, and four cases (5, 6, 10, 15) fail in all four. The assistant suspected the gold label of scr/correct/* item log-a#2 ("the models may be right and the label wrong") and left the call to Alan.
grounding 76.8 repeats qwen3-coder-30b's 76.8 to the decimal, a different model and generation. nullschema 56.9 is better than the 30B coder's 37.5.
nullschema did not predict real stub-data quality. In a DSH goal-loop task (stock-control platform), Coder produced 28,094 records with every foreign key valid, so the prediction "dangling foreign keys or malformed JSON" failed. The assistant's own conclusion: the suite measures refusal under adversarial pressure, which is not construction.
Verification behaviour followed reasoning, not the harness. Same task and goal prompt: qwen3.6 (thinking) made 24 curl self-checks; Coder made 0 and spawned two subagents instead.
Long context held. The last turn of the DSH session ran at 193,443 tokens with no truncation, 14.7 t/s.
Never measured: coinpusher (the one prediction left open), the agentic suites (tools, toolscale, chainwalk; "-- NONE --" for this label on 2026-09-07), tagger suites, contract/toolcall.

Why it was tested, in the lab's words

Alan's hunch about the previous night. 2026-08-27 07:06Z, after the Thinking result: "Qwen coder probably what I should’ve been using for the coding last night". The assistant checked the lab's own data for the 30B coder rather than assuming: qwen3-coder-30b matched Instruct's bloglanding quality but took 5x as long and scored nullschema 37.5, so "for last night specifically, you had it right".
Alan, 07:08Z: "Qwen coder next is specialised at coding no?". The assistant found it: same architecture as the model already running, non-thinking, "the best-fitting candidate I've seen for you all session". Alan: "Pull it also there are now smaller distills of 3.8 kicking about now".
Alan turned it into a controlled experiment, 07:15Z, declining the assistant's offer to add the small Qwen3.8 distills: "Nah let’s stay with the qwen next stuff for now, 3 models, same family, same quant, same context, let’s learn why they exist and what they’re best at". The assistant: "one architecture, one quant, one context window, three post-training regimes".
The behaviour it was expected to fix: the Instruct model's failed Doom-clone session (five http.server starts, wrong job killed, exit code ignored; blank canvas). The preset comment says Coder's "recovery from execution failures" training targets exactly that (fleet-models.ini, CODER block).

Open questions

Does Coder-Next win coinpusher and the agentic suites? The original reason to fetch it is untested at the board level.
Is the 76.8 grounding tie with the 30B coder a coincidence or a family trait of coding post-training?
Do the four universal scr/correct false rejects need a relabel? Alan's call; none is on record.

The model in detail

field	value
vendor	Qwen (Alibaba)
upstream repo	`Qwen/Qwen3-Coder-Next` (and `-FP8`); nothing downloaded from it. File came from `unsloth/Qwen3-Coder-Next-GGUF`
revision	not on record
licence	not on record
architecture	80B total, ~3B active, same Gated DeltaNet plus gated-attention hybrid as Qwen3-Next, 262,144 native context
reasoning	non-thinking only. The gate probe returned 0 reasoning characters where the Thinking sibling emitted 790 for the same question
vendor-published scores (claims, not measured here)	SWE-bench Verified 70.6, SWE-bench Pro 44.3, Terminal-Bench 2.0 36.2; trained for tool-calling and "recovery from execution failures"
vendor sampling	temp 1.0 / top_p 0.95 / top_k 40; the harness pins temp 0
release date	not on record. An HF listing shows `Qwen/Qwen3-Coder-Next-GGUF` dated 2026-02-02
runtime used here	llama.cpp b10618 `eb25b7263` (run log)