Ling 3.0 Flash vs Thinking: What Changed in Our Test
On four short prompts, Ling 3.0 Flash and Ling 3.0 Flash Thinking both passed our binary checks. On two prompts that required checking several constraints or counting cases, regular Flash repeated the same wrong answer three times. Thinking got both right.
In web searches on July 24, 2026, we did not locate a public Ling 3.0 Flash model card or official benchmark table. This article therefore reports a small comparison through NanoGPT rather than trying to rank the model against a wider field.
How NanoGPT lists the two versions
NanoGPT currently lists Ling 3.0 Flash as a 124-billion-parameter mixture-of-experts model with about 5.1 billion parameters active for each token. The active count describes its architecture; it does not predict answer quality, latency, or retail pricing on its own.
The two NanoGPT variants share:
- A 262,144-token input limit
- Up to 32,768 output tokens
- Tool calling
- The same listed token rates
- Inclusion in the NanoGPT subscription
Regular Flash keeps reasoning off by default. Thinking enables reasoning at medium effort and also exposes none, minimal, low, medium, high, and xhigh effort levels.
What we tested
We sent six short prompts to both variants through NanoGPT's non-streaming Chat Completions API on July 24. Each prompt was run three times at temperature 0. Regular Flash used its default reasoning setting of none; Thinking used its default of medium. Calls ran sequentially as fresh conversations, with no retries, errors, exclusions, or output-token cap. Timing covered the complete request rather than time to first streamed text. These API calls used metered billing, separate from subscription access.
The prompts covered a one-line code repair, exact JSON extraction, a dependency schedule, tool-call generation and argument formatting, a five-item ordering puzzle, and a counting problem. Answers were checked against exact expected values or, for the code task, against either of two accepted fixes. The tool prompt named the function, and that was the only tool supplied.
This is a small practical test, not a general benchmark. Repeating each prompt showed limited repeatability in this setup, but three identical runs do not turn one logic puzzle into broad evidence about reasoning ability.
Results
| Prompt | Regular passing calls | Thinking passing calls | What happened |
|---|---|---|---|
| One-line code repair | 3/3 | 3/3 | Both returned valid fixes |
| Exact JSON extraction | 3/3 | 3/3 | Both preserved every value and field |
| Dependency schedule | 3/3 | 3/3 | Both returned 16 minutes |
| Tool-call generation and arguments | 3/3 | 3/3 | Both produced the requested call correctly |
| Five-talk ordering puzzle | 0/3 | 3/3 | Regular violated one adjacency rule |
| Five-digit counting problem | 0/3 | 3/3 | Regular returned 144; the answer was 1,560 |
Regular Flash passed four of the six prompts. Thinking passed all six. The gap came entirely from the ordering and counting prompts.
Where the outcomes differed
The ordering prompt placed five talks into five slots under several rules. Regular Flash returned DBEAC in all three runs. That order satisfies most of the conditions, but places A directly beside E despite an explicit rule forbidding it. Thinking returned the unique valid order, ADBEC, every time.
The counting prompt asked how many five-digit numbers can be made from the digits 0 through 7 without repetition when the result must be divisible by 5. Regular Flash answered 144 in all three runs. Thinking returned 1560, with its reasoning separating the two possible final digits:
- Final digit 0:
7 × 6 × 5 × 4 = 840 - Final digit 5:
6 × 6 × 5 × 4 = 720 - Total:
840 + 720 = 1,560
These were not obscure knowledge questions. Both required holding several conditions in place and checking the result before answering. On these two prompts, the Thinking route passed every run in which regular Flash failed.
Where the binary outcome stayed the same
On the other four prompts, both variants passed every run.
For the code repair, regular Flash changed the initializer to -Infinity; Thinking used numbers[0]. Both are valid one-line fixes for a non-empty array. Both variants also produced the exact requested invoice JSON, calculated the 16-minute project schedule, and created the requested currency lookup tool call with the correct currencies and date.
Across all 18 requests per variant, NanoGPT reported 2,379 prompt tokens and 333 completion tokens for regular Flash. Thinking reported 2,361 prompt tokens and 6,223 completion tokens. Its final answers were still short, while the responses also contained a separate reasoning field.
NanoGPT listed both variants at $0.06 per million input tokens and $0.18 per million output tokens during the test. The full regular run cost about $0.000203, while Thinking cost about $0.001259. Thinking reported 18.7 times as many completion tokens, but total cost was 6.2 times higher because both runs also paid for roughly the same prompt-token total. The absolute amounts were tiny because the prompts were short.
Median end-to-end time was 1.50 seconds for regular Flash and 2.00 seconds for Thinking. That comparison includes the two prompts regular Flash answered quickly but incorrectly. Looking only at the four prompts both passed, the medians were 1.52 and 1.75 seconds.
Reported completion-token use also varied. Two Thinking runs of the same code-repair prompt reported 103 completion tokens, while the third reported 715, even though all three produced the same final line. Cost estimates for repeated work should leave room for that variation.
What this test supports
Regular Flash passed every run of our extraction, fixed-format code change, schedule, and tool-call prompts while using fewer reported tokens. One example from each category is not enough to establish a general capability, but it gives a reason to test the regular variant first on a representative direct prompt.
The Thinking route passed the two deduction prompts on which regular Flash failed. We did not test planning, difficult debugging, repository work, or long tool chains, so these results should not be extended to those categories without more evidence.
The Thinking variant also lets API users adjust reasoning effort rather than staying at the medium default. We did not test that full range here. A lower effort may be enough for some prompts, while harder work may justify high or xhigh.
Our test does not establish how Ling 3.0 performs on large repositories, long tool workflows, or established coding benchmarks. It shows a smaller result: regular Flash passed four individual prompts and failed two, while medium-effort Thinking passed all six at higher reported token use and cost.
Test prompts and answer keys
The tests used the following instructions, shortened here only where the requested JSON shape is already described in the label:
- Code repair: Fix one line in a maximum-value function initialized with
best = 0, for a non-empty array that may contain only negative numbers. Accepted:best = -Infinityorbest = numbers[0]. - JSON extraction: Extract invoice Z-204 with 17 adapters at EUR 8.40 and 3 cables at EUR 5.50 into the supplied exact JSON shape.
- Dependency schedule: A takes 4 minutes; B takes 7 after A; C takes 3 after A; D takes 5 after B and C; two workers are available. Expected:
16. - Tool call: Use the only supplied function,
lookup_exchange_rate, for EUR to JPY on2026-07-24. Expected arguments:base=EUR,quote=JPY,date=2026-07-24. - Ordering puzzle: Place A through E in five slots where A is before C, B is immediately after D, E is neither first nor last, C is after E, and A is not adjacent to E. Expected:
ADBEC. - Counting: Count five-digit numbers formed from 0 through 7 without repetition, divisible by 5, and not starting with 0. Expected:
1560.
Open Ling 3.0 Flash or Ling 3.0 Flash Thinking on NanoGPT.
