Models / deepseek-v4-1-flash
DeepSeek V4.1 Flash
DeepSeek's model, with open weights. Forcebench has run it in 3 configurations on the API and Local boards, plus 2 on the lite subset. API is models reached through a hosted service; Local, models served on Forcebench’s own hardware.
On the API board
Models reached through a hosted service. DeepSeek V4.1 Flash is ranked by its one configuration: the 12th-highest score of 26 models, not clearly different from 6 other models. Scores are within 3 attempts (c@3) with 95% intervals; each configuration is compared with the highest-scoring one by whether their 95% intervals overlap. Every configuration on the board → Cost and speed →
| Configuration | Rank all configs | Within 3 attempts c@3 · 95% CI | Tokens mean · median · p90 | No answer | Latency median per task, every attempt counted | Cost per task | Cost per solved task |
|---|---|---|---|---|---|---|---|
| OpenRouter · FP8 · highhighest score | 20 | 81%95% CI 77–85 | 9.6k6.8k · 27k | 0% | 136 s | $0.007 | $0.0086 |
- 95% interval
- API board median
Show the numbers
| Suite | c@3 95% CI | API median | Tasks |
|---|---|---|---|
| Apex | 95%95% CI 76–99 | 90% | 20 |
| Salesforce APIs | 83%95% CI 61–94 | 72% | 18 |
| CI/CD | 88%95% CI 66–97 | 88% | 17 |
| Salesforce CLI | 95%95% CI 76–99 | 90% | 20 |
| Salesforce docs | 29%95% CI 13–53 | 29% | 17 |
| fflib Enterprise Patterns | 67%95% CI 44–84 | 61% | 18 |
| Flow | 61%95% CI 39–80 | 69% | 18 |
| Governor limits & pushback | 89%95% CI 69–97 | 89% | 19 |
| Lightning Web Components | 100%95% CI 82–100 | 100% | 18 |
| Nonprofit Success Pack | 89%95% CI 67–97 | 81% | 18 |
| Packaging | 78%95% CI 55–91 | 69% | 18 |
| Permissions & access | 87%95% CI 62–96 | 87% | 15 |
| Scratch org definitions | 61%95% CI 39–80 | 50% | 18 |
| SOQL | 95%95% CI 76–99 | 100% | 20 |
| Trigger Actions Framework | 94%95% CI 74–99 | 86% | 18 |
On the Local board
Models served on Forcebench’s own hardware. DeepSeek V4.1 Flash is ranked by the highest-scoring of its 2 configurations: the highest score of 9 models, not clearly different from 1 other model. Scores are within 3 attempts (c@3) with 95% intervals; each configuration is compared with the highest-scoring one by whether their 95% intervals overlap. Every configuration on the board →
- 95% interval
Show the numbers
| Configuration | Rank all configs | Within 3 attempts c@3 · 95% CI | Against the highest by 95% interval overlap | Tokens mean · median · p90 | No answer | Latency median per task, every attempt counted |
|---|---|---|---|---|---|---|
| EXL3 2.9bpw · vLLM + ExLlamaV3 · high · 2× NVIDIA DGX Spark (GB10, 128 GB each)highest score | 1 | 73%95% CI 68–78 | — | 7.4k3.5k · 22k | 6% | 161 s |
| Native FP8, FP4 experts · SGLang · high · 3× NVIDIA GB10, 128 GB each (2× DGX Spark, 1× Gigabyte AI TOP ATOM) | 2 | 70%95% CI 66–75 | not clearly different | 9.6k6.1k · 26k | 8% | 253 s |
Latency depends on the hardware: compare it only between configurations on the same hardware.
- 95% interval
- Local board median
Show the numbers
| Suite | c@3 95% CI | Local median | Tasks |
|---|---|---|---|
| Apex | 95%95% CI 76–99 | 70% | 20 |
| Salesforce APIs | 78%95% CI 55–91 | 44% | 18 |
| CI/CD | 82%95% CI 59–94 | 53% | 17 |
| Salesforce CLI | 85%95% CI 64–95 | 45% | 20 |
| Salesforce docs | 24%95% CI 10–47 | 12% | 17 |
| fflib Enterprise Patterns | 50%95% CI 29–71 | 17% | 18 |
| Flow | 61%95% CI 39–80 | 17% | 18 |
| Governor limits & pushback | 89%95% CI 69–97 | 84% | 19 |
| Lightning Web Components | 89%95% CI 67–97 | 89% | 18 |
| Nonprofit Success Pack | 83%95% CI 61–94 | 61% | 18 |
| Packaging | 44%95% CI 25–66 | 39% | 18 |
| Permissions & access | 80%95% CI 55–93 | 60% | 15 |
| Scratch org definitions | 44%95% CI 25–66 | 17% | 18 |
| SOQL | 95%95% CI 76–99 | 90% | 20 |
| Trigger Actions Framework | 94%95% CI 74–99 | 44% | 18 |
By quantisation
The same effort in different quantisations, each against the highest-precision one on the same tasks (a paired 95% interval), on pass@1, the first answer alone. Where the engine changes too, a difference can't be put down to quantisation alone. The quantisation study has every model.
- 95% interval
Show the numbers
| Build | pass@1 | 95% CI | Tokens |
|---|---|---|---|
| Native FP8, FP4 experts · SGLang · 3× NVIDIA GB10, 128 GB each (2× DGX Spark, 1× Gigabyte AI TOP ATOM) | 57% | 52–62 | 9.6k |
| EXL3 2.9bpw · vLLM + ExLlamaV3 · 2× NVIDIA DGX Spark (GB10, 128 GB each) | 59% | 54–64 | 7.4k |
| Quantisation · engine · high | pass@1 95% CI | Against Native FP8, FP4 experts same tasks, paired | Tokens |
|---|---|---|---|
| Native FP8, FP4 experts · SGLang · 3× NVIDIA GB10, 128 GB each (2× DGX Spark, 1× Gigabyte AI TOP ATOM)reference | 57%95% CI 52–62 | — | 9.6k |
| EXL3 2.9bpw · vLLM + ExLlamaV3 · 2× NVIDIA DGX Spark (GB10, 128 GB each) | 59%95% CI 54–64 | no clear difference (+2 pts, 95% CI −3 pts to +7 pts); this data can't rule out a change of up to 7 points | 7.4k |
In the studies
- Max effort: what its max effort did against the effort below it.
- Quantisation: its quantisations side by side.
- Local vs API: where it lands among the configurations of both boards.
- Pushback: whether it pushed back on requests that break at scale.
- Coding agent: the same tasks answered by a coding agent, with and without skills.
Private evaluations
Want this for your team?
Azul Labs runs Forcebench privately for two kinds of Salesforce team.
Choosing a model to run yourself?
A private benchmark of your shortlist
Open-weight models at the quantisations and reasoning efforts you could run on infrastructure you control, tested on your kind of Salesforce work, with a written recommendation.
- Private tasks written from your work only ever run on local models; they are never sent to an API.
- Re-testing after any change of model, quantisation or engine is available.
- Typical turnaround is 2–3 weeks; large models take about a day of machine time per configuration.
Building or buying a Salesforce AI tool?
A private evaluation of your tool
Your tool on held-out tasks it has never seen, against API baselines (models reached through a hosted service) and Local ones (served on Forcebench’s own hardware), with a written report on where it holds up and where it fails.
- It can start as a smaller pilot, about 30 held-out tasks with your tool against two baselines, before a full evaluation.
- Repeat evaluations after your releases are available.
- Typical turnaround is 2–3 weeks.
Fixed fee, quoted up front. The quote depends on the number of tasks and the number of models, tools or configurations tested.
Or email leo@azl.au · azl.au